[mlir][bufferize] Infer memref types when possible

Instead of recomputing memref types from tensor types, try to infer them when possible. This results in more precise layout maps.

Differential Revision: https://reviews.llvm.org/D125614