[DEBUG] test all float dtypes for images, videos, and boxes

[pmeier]

No description provided.

[pmeier]

After I freaked a little last night seeing 4k failing tests, @vfdev-5 rightfully pointed out that most of v1 does not support float16 either. This is why @NicolasHug only saw failures for ElasticTransform in #7159 rather than for all transforms that use _apply_grid_transform internally. It is an outlier for two reasons:

1. The computation is simple enough that the only op that doesn't support float16 is grid_sample.

2. For the reason above, we probably decided to test it explicitly:
   

   vision/test/test_functional_tensor.py

   Line 1285 in 5ea8e01

   @pytest.mark.parametrize("dt", [None, torch.float32, torch.float64, torch.float16])

   In contrast, the tests for F.affine have this escape hatch:

   vision/test/test_functional_tensor.py

   Lines 159 to 161 in 5ea8e01

   	if dt == torch.float16 and device == "cpu":
   	# skip float16 on CPU case
   	return

[vfdev-5]

Suggested change

	sizes=["random"], dtypes=[torch.uint8, torch.float16, torch.float32, torch.float64]
	sizes=["random"], dtypes=[torch.uint8, torch.float32, torch.float64]

Shouldn't we remove f16 from here as well ?

[pmeier]

Handled in #7211.