Enable autocast for all ops

[fmassa]

🚀 Feature

Currently only NMS and RoIAlign support autocast (introduced in PyTorch 1.6.0). We should extend support for autocast to all C++ ops in torchvision.

Motivation

Mixed precision training is PyTorch is obtained via support for autocast (automatically casting between fp16 -> fp32 depending on the operation).

The work should follow what was done for the RoIAlign function in #2384 (as that PR also converted NMS to use the dispatcher).

Adding support for autocast means that the op should be registered via the dispatcher, so this is blocked by #2796

Pitch

Out-of-the-box mixed precision support for all operators implemented in torchvision:

• nms
• roi_align
• deform_conv2d
• roi_pool
• ps_roi_align
• ps_roi_pool

[fmassa]

Fixed via referenced issues, with last PR in #2938