[Operator] Support ndim > 3 for batch_dot

[sxjscience]

In batch_dot, i.e, C = batch_dot(A, B), both inputs A & B must have ndim=3. However, in some cases our inputs could have have higher dimensionality. We can consider to support the following two cases:

• Both A, B have ndim >3

  We should compute the output as:

  C[..., i, j] = sum_k (A[..., i, k] * B[..., k, j]), for all indices i, j.
  

  For example, if the shape of A is (batch_size, a1, a2, …, b, c) and the shape of B is (batch_size, a1, a2, …, c, d), the shape of C will be (batch_size, a1, a2, ..., b, d).
  This is similar to the matmul in Tensorflow

• One of A, B have ndim > 3

  Assume A.ndim >3, we have

  C[b, ..., i, j] = sum_k (A[b, ..., i, k] * B[b, k, j]), for all indices i, j.
  

Supporting these two cases will solve the problems in https://discuss.gluon.ai/t/topic/2534/8, https://discuss.gluon.ai/t/topic/5618/9

[tqchen]

maybe related to optimizations in http://www.tvmlang.org/2018/03/23/nmt-transformer-optimize.html

[yufengwhy]

这个issue解决了吗？我看mxnet 1.2里面没有提到这个issue？

[sxjscience]

I'm working on it.

[Angzz]

同问，1.3 pre版本的都已经更新了，但是貌似还是只支持3 dim.

[yufengwhy]

数据量较大，reshape有点慢，弄好了通知一声，更新一下代码，@sxjscience