Ffcv improvements

[lrzpellegrini]

Minor FFCV support improvements:

• Add CenterCrop transformation to support variable-resolution datasets when loaded with CenterCrop as the first transform (a decoder that achieves something similar already exists in FFCV, but its behavior is not aligned with the torchvision one)
• Fix the FFCV try_speed example
• Add non_blocking=True in _unpack_minibatch of supervised problem and object detection

[coveralls]

Pull Request Test Coverage Report for Build 6484704365

• 1 of 1 (100.0%) changed or added relevant line in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at 72.5%

---

Totals
Change from base Build 6469972036:	0.0%
Covered Lines:	17566
Relevant Lines:	24229

---

💛 - Coveralls

[AntonioCarta]

looks good. You can merge it yourself once you fix the lint issue ;)