Add batch/microbatch transforms

[mvpatel2000]

What does this PR do?

#3566 moved device transfers into the microbatch loop. This results in device transformations occurring on CPU, which can be slow as pointed out in #3699.

To remedy this, we deprecate device_transforms (which was poorly named anyways) and instead add batch_transforms and microbatch_transforms to dataspec. Batch transforms happens on CPU, and microbatch transforms happens on GPU after the device transfer.

Additionally, if users want to do batch level transforms on GPU, they can add a transform that moves to device.

[mvpatel2000]

CC: @Ghelfi

[antoinebrl]

Thanks @mvpatel2000 ! Will there be a new release following this merge? The regression introduced in 0.25.0 is a blocker for us as our trainings are 10x slower.

[mvpatel2000]

Thanks @mvpatel2000 ! Will there be a new release following this merge? The regression introduced in 0.25.0 is a blocker for us as our trainings are 10x slower.

We're still figuring out when it would be released. To unblock, I recommend adding a transform that just moves batch to device.

[antoinebrl]

To unblock, I recommend adding a transform that just moves batch to device.

Thanks! That is exactly the patch we have right now 👍

[antoinebrl]

LGTM! Thanks

[Ghelfi]

Thanks for quick implementation.
@mvpatel2000 given the fact that we cannot, for now, run any transform on device, it seems appropriate to trigger a release o re-enable this functionality since it is now patched.