[LLVMCPU] Refactor multi lowering config propagation and setting (iree-org#22126)

The previous propagation logic was fragile because it assumed all ops in
a dispatch had an identity mapping between their parallel dimensions. It
also relied on indexing with `maxLoopNums`, which breaks for cases like
mmt4d.

This change replaces that approach with `IterationDimTracker +
totalLoopNum` for more robust propagation.

Additionally, a new helper class `MultiLoweringConfigGenerator` is
introduced to streamline the configuration process.

Closes: iree-org#21696

---------

Signed-off-by: Yu-Zhewen <[email protected]>

Author: Yu-Zhewen