Enable multi-dimensional and axis support for tf.unique_with_counts

This fix tries to address the issue raised in 16499 to bring
multi-dimensional and axis support for `unique_with_counts`.

When `UniqueV2` kernel was added in 12952, it actually supports
multi-dimensional and axis support for `unique_with_counts` as well,
just not registered.

This fix:
1. Register `UniqueWithCountsV2` kernel to have axis support.
2. Hide both `UniqueWithCounts` and `UniqueWithCountsV2`
3. Add python unique_with_counts wrapper to call `gen_array_ops._unique_with_counts`
4. If APi review passes and the PR merges, `unique_with_counts` will switch to `gen_array_ops._unique_with_counts_v2` (in 3 weeks).

This fix fixes 16499.

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