Converting tosa.reshape to tensor.expand_shape and
tensor.collapse_shape logically belongs in the tosa-to-tensor
conversion process. In addition, we (rocMLIR downstream) want to use
the reshape -> expand/collapse_shape logic to simplify parts of our
Tosa integration without using the full tosa-to-linalg flow, further
motivating moving these patterns.
The downside to this change is that it means you need to run
tosa-to-tensor after tosa-to-linalg, which is probably a breaking
change.