[MLIR] Loop invariant cache generalization

we should generalize this for the lower-upper bound check instead of just alloca/memref load (incl affine load).

Start by locating the MLIR loop-invariant cache logic that handles alloca and memref loads, including affine loads. Trace how the lower-upper bound check is currently handled and identify the relevant tests; done means the check is generalized beyond those load cases.

Written by the indexing model from the issue text.