[NPU] Add NPU optimized mHC

[noemotiovon]

Add Ascend NPU Triton kernels for the three mHC sub-operators:

• Fused matmul + RMS normalization (forward/backward)
• Sinkhorn routing with split pre/post/residual coefficients (forward/backward)
• Pre-aggregate weighted sum (forward/backward)
• Post + residual mixing (forward/backward)

NPU optimizations applied:

• Unified UB tiling via compute_default_tiling_strategy
• Persistent grid-stride loops (tl.range + num_programs)
• Cache modifiers (.cg/.ca/.cs) on all loads/stores
• Adaptive BLOCK_N/BLOCK_M for core utilisation at small seq_len
• Fused backward coefficient assembly kernel (replaces ~20 PyTorch ops with 1 Triton kernel)

Hardware Type: Atlas 800I A2

• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

[noemotiovon]

Hi @lowdy1, could you continue my work?

[lowdy1]

Hi @lowdy1, could you continue my work?

If you don’t have the bandwidth right now, I'm glad to continue working on this.

[noemotiovon]

Hi @lowdy1, could you continue my work?

If you don’t have the bandwidth right now, I'm glad to continue working on this.

Thank you!