MT model supports UpsamplerPixelShuffle 4x upsampling

Summary:
MT supports efficient 4x upsampling using UpsamplerPixelShuffle, with main changes below
- **UpsampleHeadMask**: support **subpixel_stride**. when it is 4, use HTP efficient **UpsamplerPixelShuffle**.
- eval data loader mt_eval_vga.yaml: set num_max_samples = 20K to reduce running time of async eval op. Otherwise, it will be eliminated due to low gpu util
- pixel_shuffle_use_conv_transpose: support to simulate F.pixel_shuffle using a 2D convtranspose op. But on G2, we find it is slower since it will run on HMX (vs HVX), which already has heavyweight workload from various 2d conv ops

Reviewed By: hyxu2006

Differential Revision: D77117618

fbshipit-source-id: e0a39526e558e3ddb9ee75d953ed28ffde7ba722

Author: stephenyan1231

detectron2/utils/visualizer.py

@@ -1242,7 +1242,7 @@ def _convert_masks(self, masks_or_polygons):
             m = m.numpy()
         ret = []
         for x in m:
-            if isinstance(x, GenericMask):
+            if isinstance(x, GenericMask) or x.__class__.__name__ == "GenericMask":
                 ret.append(x)
             else:
                 ret.append(GenericMask(x, self.output.height, self.output.width))