[Codegen] Add pass to further tile large linalg ops (#19010)

In certain cases/configurations we can end up with excessively large
static `linalg.generic` ops that will then vectorize and emit code with
huge vector sizes. This at a minimum will yield terrible code, and can
even crash/hang LLVM when trying to compile such an enormous function.

The implementation here only looks at parallel dims and linalg ops.
Large reductions and other vectorizable ops will likely need handling as
well, but "vectorizability" isn't really queryable today.

The pass has an options for a maximum vector size to tile to, but the
logic here could potentially use some future improvement (e.g. maybe
expose a target vector size too to use when we choose to tile).

Also note that in many cases where we hit this is likely to generate
poor code because we can often end up with large local/stack
allocations, but by the time we get to this point there isn't much else
we can do. Better to ask for an allocation than 10k lines of unrolled
IR.