NewCclCommMgr on runtime

[ShawnXuan]

问题描述：
import oneflow时会创建一些全局对象，其中下面段落创建了NewCclCommMgr对象。
https://github.com/Oneflow-Inc/oneflow/blob/master/oneflow/core/job/env_global_objects_scope.cpp#L200-L208

不过当使用非CUDA设备时，无法创建NewCclCommMgr对象，在runtime阶段带来一些问题。

这里的改动，就是在runtime阶段也尝试创建NewCclCommMgr对象。

注：NewCclCommMgr对象的创建，依赖RegisterEagerCclCommMgrType

没有解决的问题：在runtime之前，NewCclCommMgr对象仅仅用于InsertNcclLogicalOpPass，我觉得，并不是必须的，或许就可以把NewCclCommMgr对象的创建放到后面，不过涉及的范围比较大，怕有其他副作用。

[github-actions]

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

[github-actions]

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10617/

[github-actions]

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.2ms (= 4320.0ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.1ms (= 5710.9ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.32 (= 57.1ms / 43.2ms)

OneFlow resnet50 time: 26.2ms (= 2617.8ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 38.3ms (= 3828.9ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.46 (= 38.3ms / 26.2ms)

OneFlow resnet50 time: 19.1ms (= 3828.8ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.3ms (= 7063.8ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.84 (= 35.3ms / 19.1ms)

OneFlow resnet50 time: 17.5ms (= 3497.3ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 32.0ms (= 6403.8ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.83 (= 32.0ms / 17.5ms)

OneFlow resnet50 time: 17.5ms (= 3496.4ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.1ms (= 5825.7ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.67 (= 29.1ms / 17.5ms)

OneFlow swin dataloader time: 0.199s (= 39.752s / 200, num_workers=1)
PyTorch swin dataloader time: 0.127s (= 25.410s / 200, num_workers=1)
Relative speed: 0.639 (= 0.127s / 0.199s)

OneFlow swin dataloader time: 0.054s (= 10.836s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.550s / 200, num_workers=4)
Relative speed: 0.604 (= 0.033s / 0.054s)

OneFlow swin dataloader time: 0.031s (= 6.212s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.309s / 200, num_workers=8)
Relative speed: 0.533 (= 0.017s / 0.031s)

❌ OneFlow resnet50 time: 49.4ms (= 4935.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 63.8ms (= 6382.6ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.29 (= 63.8ms / 49.4ms)

OneFlow resnet50 time: 36.8ms (= 3684.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 46.2ms (= 4615.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.25 (= 46.2ms / 36.8ms)

OneFlow resnet50 time: 27.7ms (= 5530.4ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 39.3ms (= 7865.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.42 (= 39.3ms / 27.7ms)

OneFlow resnet50 time: 25.5ms (= 5090.5ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 39.4ms (= 7878.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.55 (= 39.4ms / 25.5ms)

OneFlow resnet50 time: 25.0ms (= 4994.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 35.9ms (= 7179.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.44 (= 35.9ms / 25.0ms)

[github-actions]

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10617/

[github-actions]

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.6ms (= 4360.3ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.4ms (= 5745.0ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.32 (= 57.4ms / 43.6ms)

OneFlow resnet50 time: 26.2ms (= 2617.6ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 38.0ms (= 3802.0ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.45 (= 38.0ms / 26.2ms)

OneFlow resnet50 time: 19.0ms (= 3799.8ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.2ms (= 7047.0ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.85 (= 35.2ms / 19.0ms)

OneFlow resnet50 time: 17.6ms (= 3510.9ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 32.3ms (= 6460.7ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.84 (= 32.3ms / 17.6ms)

OneFlow resnet50 time: 16.9ms (= 3371.1ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 28.1ms (= 5625.4ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.67 (= 28.1ms / 16.9ms)

OneFlow swin dataloader time: 0.199s (= 39.899s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.760s / 200, num_workers=1)
Relative speed: 0.646 (= 0.129s / 0.199s)

OneFlow swin dataloader time: 0.057s (= 11.499s / 200, num_workers=4)
PyTorch swin dataloader time: 0.032s (= 6.497s / 200, num_workers=4)
Relative speed: 0.565 (= 0.032s / 0.057s)

OneFlow swin dataloader time: 0.031s (= 6.111s / 200, num_workers=8)
PyTorch swin dataloader time: 0.016s (= 3.280s / 200, num_workers=8)
Relative speed: 0.537 (= 0.016s / 0.031s)

❌ OneFlow resnet50 time: 49.6ms (= 4962.9ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.2ms (= 6515.9ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.31 (= 65.2ms / 49.6ms)

OneFlow resnet50 time: 35.6ms (= 3562.6ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 44.6ms (= 4456.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.25 (= 44.6ms / 35.6ms)

OneFlow resnet50 time: 28.0ms (= 5606.4ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 42.5ms (= 8501.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.52 (= 42.5ms / 28.0ms)

OneFlow resnet50 time: 25.4ms (= 5089.3ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 39.2ms (= 7831.9ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.54 (= 39.2ms / 25.4ms)

OneFlow resnet50 time: 24.8ms (= 4965.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.0ms (= 7198.7ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.45 (= 36.0ms / 24.8ms)