Enable half and double for resampler GPU ops

[WindQAQ]

Actually, I don't know how we passed tests previously...
We don't register half and double backward kernel, and even don't specialize half forward kernel. But we indeed test them.

https://github.com/tensorflow/addons/blob/master/tensorflow_addons/image/tests/resampler_ops_test.py#L95-L124
https://github.com/tensorflow/addons/blob/master/tensorflow_addons/image/tests/resampler_ops_test.py#L185-L213
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/util/gpu_device_functions.h#L670-L691

[seanpmorgan]

This is puzzling to me. I checked on previous kokoro logs and indeed test_op_backward_pass[gpu-float16] was running and passing.

TFA 0.10 sucessfully performs this as well:

print(data_tensor.dtype)
print(warp_tensor.dtype)
with tf.GradientTape() as g:
	g.watch(data_tensor)
	out = resampler_ops.resampler(data_tensor, warp_tensor)

grad = g.gradient(out, data_tensor)
print(grad[0][:1])
print(out.dtype)
print(grad.dtype)

<dtype: 'float16'>
<dtype: 'float16'>
2020-05-18 21:10:02.200555: I tensorflow/core/common_runtime/placer.cc:114] gradients_resampler_addons_resampler_grad_addons_resamplergrad_0_RetVal: (_Retval): /job:localhost/replica:0/task:0/device:GPU:0
tf.Tensor(
[[[0.4368 0.4368 0.4368]
  [0.9355 0.9355 0.9355]
  [0.     0.     0.    ]
  [0.     0.     0.    ]
  [0.     0.     0.    ]]], shape=(1, 5, 3), dtype=float16)
<dtype: 'float16'>
<dtype: 'float16'>

Regardless the change LGTM... Thanks @WindQAQ!

[seanpmorgan]

@alextp No urgency, but whenever time allows could you or someone you could direct us to take a look at why this was working before. Would like to improve our understanding of the registering mechanism.

Colab link:
https://colab.research.google.com/drive/1o53EPNALyTRFlJGLLZmihlSUfCFDHKUz?usp=sharing