Float16 (half or Eigen::half) for conv3d ops (#12832)

* fp16 on conv3d, conv3d_transpose, tf.layers.batch_norm and tf.slim.batch_norm (batchnorms without fused_batch_norm)

* fp16 for conv3d_transposed

* remove batch_norm fp16 and rewrite conv3d kernels registrations

* remove additional logging prints

* tests for conv3d gradients

* fp64 test for cpu

* input normalization and requests

* fix initialization comment and remove unnecessary variable

* lowered tolerance for fp16 test

* remove unnecessary part of the comment