[mlir] Make tensor_to_memref op docs match reality

The previous code defined it as allocating a new memref for its result.
However, this is not how it is treated by the dialect conversion framework,
that does the equivalent of inserting and folding it away internally
(even independent of any canonicalization patterns that we have
defined).

The semantics as they were previously written were also very
constraining: Nontrivial analysis is needed to prove that the new
allocation isn't needed for correctness (e.g. to avoid aliasing).
By removing those semantics, we avoid losing that information.

Differential Revision: https://reviews.llvm.org/D91382