Add GPU and CPU implementation of `tf.histogram_fixed_width`. (#13731)

* Add GPU and CPU implementation of `tf.histogram_fixed_width`.

This fix adds the GPU and CPU implementation of `tf.histogram_fixed_width`.
The previous implementation was done in python. This fix adds
C++ kernel for GPU and CPU>

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Update python ops for `tf.histogram_fixed_width`

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Update test cases to invoke GPU test for `tf.histogram_fixed_width`

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Disable int64 output on GPU for now as atomicAdd is not supported yet.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Address review feedback and use a stable version of summation.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Maintain backward compatibility of the API

Keep `dtype` and make sure `nbins = 100` is in attr

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>