Add half-precision support for `DepthwiseConv2dNative` (#12976)

* Add float16 support for `DepthwiseConv2dNative`
This fix tries to address the request raised in 12327
so that float16 is supported for `DepthwiseConv2dNative`

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

* Add test cases for float16 support of `DepthwiseConv2dNative`

A couple of tests for float16 have been disabled as error is
large.

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

* Add float16 support for `DepthwiseConv2dNative` on GPU

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

Add Eigen::half support for DepthwiseConv2dNative on GPU

This commit add Eigen::half support for DepthwiseConv2dNative on GPU

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

* Disable float16 support for the gradient of DepthwiseConv2dNative

The accuracy is not high enough for float16 of gradient.

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