upsample_bilinear: fix output data-type.

[ysiraichi]

Fix: #7095

This PR removes the data-type conversion inherited from the TF implementation when is_kernel_bilinear is set. We still do it on integer data-types, though.

Note: upsample_bilinear call on integral tensors (except torch.byte type) is not currently implemented on upstream PyTorch. It fails with:

Traceback (most recent call last):
  File "examples/upsample.py", line 16, in <module>
    out = foo(a)
  File "examples/upsample.py", line 7, in foo
    return torch.nn.functional.upsample_bilinear(x, scale_factor=2)
  File "torch/nn/functional.py", line 4203, in upsample_bilinear
    return interpolate(input, size, scale_factor, mode="bilinear", align_corners=True)
  File "torch/nn/functional.py", line 4082, in interpolate
    return torch._C._nn.upsample_bilinear2d(input, output_size, align_corners, scale_factors)
RuntimeError: "upsample_bilinear2d_channels_last" not implemented for 'Long'

cc @miladm @JackCaoG

[JackCaoG]

Seems like you need to resolve the conflict first.

[JackCaoG]

I am going to revert this pr as it breaks the TPU CI. considering makeing this change GPU only.