[mlir][tensor][bufferize] Lower tensor.generate to linalg.map

There is no memref equivalent of tensor.generate. The purpose of this change is to avoid creating scf.parallel loops during bufferization.

Differential Revision: https://reviews.llvm.org/D136767