Blockwise inplace optimizer does not filter out duplicated/aliased inputs from candidates

[ricardoV94]

Description

An inplace rewrite that tries to destroy the equivalent inputs multiple times will always be rejected by PyTensor (except for a very rare case where the op has a special property). This means Blockwise should discard such candidates, and only try remaining ones.

This can be observed with the new tridiagonal lu factor Op from #1382, but should be tested with a standalone Op in case that op is later removed.