Support for torch.float16 to half VRAM requirements

[thekitchenscientist]

If you add the code:
if latents_dtype not in [torch.float, torch.double]: noise_guidance_edit_tmp = noise_guidance_edit_tmp.float()
before the torch.quantile() calculation and then

if latents_dtype not in [torch.float, torch.double]: noise_pred = noise_pred.to(torch.float16)
prior to # compute the previous noisy sample x_t -> x_t-1

the model runs with half the VRAM, which with xformers/attention slicing enabled on the pipe means the model can run with just 5-6GB VRAM