Decrease register usage for Xe2 MoE GEMM

[sanchitintel]

Description

This change is needed for the PR that adds the MXFP4/MXFP8 MoE GEMM example.
Xe2 doesn't natively support these datatypes, so the weights are converted to FP16/BF16, and scales are converted to FP32 (technically, we can keep them in FP16/BF16, and then do elementwise multiplication while applying scales, but that entails a bit more compute), thereby increasing register pressure. This PR decreases register pressure to boost MXFP4/MXFP8 MoE GEMM performance (which use features from a couple of other existing PRs. Will update description shortly).

MoE GEMM computation has 3 components -

1. Tiling schemes (can also help decrease register pressure)
2. vanilla GEMM kernel (for MXFP4/MXFP8, it's the corresponding scaledMM)
3. tile scheduler for various output blocks/tiles --> this PR modifies it, and also assumes that igc will optimize out some unused data structures, and they won't take up unnecessary register space during GEMM computation.

Type

• Bug - [ ] Feature - [x] Performance - [ ] Refactor

Testing

• Tests pass - [ ] Xe12 - [ ] Xe20

Performance

measured with MXFP4/MXFP8 scaledMM

Metric	Before	After

References

Fixes #

Checklist

• Copyright - [ ] Co-pilot Review - [ ] Deprecated APIs not used

[sanchitintel]

Will combine several changes in one PR