[MLIR][Shape] Add conversion for missing ops to standard

This adds conversions for const_size and to_extent_tensor. Also, cast-like operations are now folded away if the source and target types are the same.

Differential Revision: https://reviews.llvm.org/D84745