refactor(mxfp4-moe): consolidate layout-API gemm into utils/moegemm/moe_dispatcher

Collapse the 5 partly-duplicated MXFP4 MoE layout-API modules into 3:

  * utils.py         -- basics (shape constants, K-size formulas, pointer/LDS
                        helpers, e8m0/SwiGLU quant math); no mma/load
  * moegemm.py       -- gemm1 + gemm2 device bodies in one file, plus the shared
                        B/B-scale fx.copy + fx.gemm layout primitives, the A-LDS
                        loaders, and the atomic bf16 epilogue
  * moe_dispatcher.py-- compile_gemm1/2_a4w4_port + launch dispatch
                        (mxfp4_moe_gemm1/2, caches, gemm1_grid)

Removes the helper duplication (gemm1 previously re-defined _raw/_lds_ptr3/
_silu_mul_batch/... already in common) and makes the stack self-contained (no
imports from the v1 modules). Pure code-motion -- no behavioral change.

Deletes mxfp4_moe_{common,layout,gemm1,gemm2,gemm_2stage}.py and rewires the
test import (kernels.mxfp4_moe_gemm_2stage -> kernels.moe_dispatcher).

Validated on gfx950: test_moe_gemm_2stage layout-API path (gemm1->gemm2,
opus-sort), a4w4 + a8w4, S/M/L -- numeric gate (mean_row_cos > 0.85) passes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: coderfeli