Resize relies on interpolate's native uint8 handling

Description:
- Now that pytorch/pytorch#90771 is merged, let Resize() rely on interpolate()'s native uint8 handling instead of converting to and from float.

  - uint8 input is not casted to f32 for nearest mode and bilinear mode if the latter has AVX2.

Context: pytorch#7217

Benchmarks:
```
[----------- Resize cpu torch.uint8 InterpolationMode.NEAREST -----------]
                         |  resize v2  |  resize stable  |  resize nightly
1 threads: ---------------------------------------------------------------
      (3, 400, 400)      |      457    |        461      |        480
      (16, 3, 400, 400)  |     6870    |       6850      |      10100

Times are in microseconds (us).

[---------- Resize cpu torch.uint8 InterpolationMode.BILINEAR -----------]
                         |  resize v2  |  resize stable  |  resize nightly
1 threads: ---------------------------------------------------------------
      (3, 400, 400)      |      326    |        329      |        844
      (16, 3, 400, 400)  |     4380    |       4390      |      14800

Times are in microseconds (us).
```

[Source](https://gist.github.com/vfdev-5/a2e30ed50b5996807c9b09d5d33d8bc2)

Author: vfdev-5

torchvision/transforms/_functional_tensor.py

@@ -459,7 +459,17 @@ def resize(
         # now we don't as True is the default.
         antialias = False
 
-    img, need_cast, need_squeeze, out_dtype = _cast_squeeze_in(img, [torch.float32, torch.float64])
+    acceptable_dtypes = [torch.float32, torch.float64]
+    if interpolation in ["nearest", "nearest-exact"]:
+        # uint8 dtype can be included for cpu and cuda input if nearest mode
+        acceptable_dtypes.append(torch.uint8)
+    elif interpolation == "bilinear" and img.device.type == "cpu":
+        # uint8 dtype support for bilinear mode is limited to cpu and
+        # according to our benchmarks non-AVX CPUs should prefer u8->f32->interpolate->u8 path
+        if "AVX2" in torch.backends.cpu.get_cpu_capability():
+            acceptable_dtypes.append(torch.uint8)
+
+    img, need_cast, need_squeeze, out_dtype = _cast_squeeze_in(img, acceptable_dtypes)
 
     # Define align_corners to avoid warnings
     align_corners = False if interpolation in ["bilinear", "bicubic"] else None

torchvision/transforms/v2/functional/_geometry.py

@@ -185,7 +185,17 @@ def resize_image_tensor(
         image = image.reshape(-1, num_channels, old_height, old_width)
 
         dtype = image.dtype
-        need_cast = dtype not in (torch.float32, torch.float64)
+        acceptable_dtypes = [torch.float32, torch.float64]
+        if interpolation in [InterpolationMode.NEAREST, InterpolationMode.NEAREST_EXACT]:
+            # uint8 dtype can be included for cpu and cuda input if nearest mode
+            acceptable_dtypes.append(torch.uint8)
+        elif interpolation == InterpolationMode.BILINEAR and image.device.type == "cpu":
+            # uint8 dtype support for bilinear mode is limited to cpu and
+            # according to our benchmarks non-AVX CPUs should prefer u8->f32->interpolate->u8 path
+            if "AVX2" in torch.backends.cpu.get_cpu_capability():
+                acceptable_dtypes.append(torch.uint8)
+
+        need_cast = dtype not in acceptable_dtypes
         if need_cast:
             image = image.to(dtype=torch.float32)