[VectorDistribute] Consider all compute ops for thread tile size (#23394)

So far, selection of the thread tile size (`threadLoads`) in reduction
VectorDistribute lowering strategy selection was purely based on the
root operation. If the selected size wasn't suitable for other compute
operations in the same dispatch, selection of the VectorDistribute
strategy would fail, falling back to other strategies.

This change addresses an existing TODO to consider the constraints of
the other compute ops in the same dispatch when choosing the thread tile
size. This may lead to smaller tile sizes (worst case one element per
thread) if the other compute ops have additional constraints, but allows
to use the VectorDistribute pipeline for such reductions.

This fixes https://github.com/iree-org/iree/issues/23340.

Assisted-by: Claude Code

ci-extra: test_torch

---------

Signed-off-by: Lukas Sommer <lukas.sommer@amd.com>