Add a python binding for xla::DotGeneral, the motivation is to represent the semantic of matmul/einsum(dtype1, dtype2)->dtype3, in which dtype3 can be configurable. Now torch.matmul would always use the inferred output dtype.

Function signature is the same as jax.lax.dot_general.

The only limitation is precision config is not supported, it can be extended easily if needed in the future.

Test:
Added unit tests