Batched multiplication support for ndims > 3

Currently NNlib.batchmul works with arrays with upto 3 dimensions. It would be nice if this could be upgraded to function similar to `numpy.matmul` or `torch.matmul` - this would help in a lot of models, especially some of the attention-based ones I'm working on 😅

Start by locating the implementation and tests for the NNlib.batchmul entry point. Compare its current dimensionality behavior with numpy.matmul and torch.matmul, then define the expected behavior for arrays above three dimensions. Done means batchmul supports the requested higher-dimensional cases and has coverage for the new semantics.

Written by the indexing model from the issue text.