Support qmatmul with different dims tensors (#4438)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4438

MobileBERT exposes an issue in our kernel, where tensors have compatible (for PyTorch) but different batch dimensions.

This diff changes the meta kernel to support that (the kernel can already do it).

Reviewed By: dulinriley

Differential Revision: D60314979

fbshipit-source-id: a0cde9d328098992787c353611ece64223d6c739