Group fixed-route MoE tokens and reuse expert weight tiles

[Happymic]

What changed

• Add an opt-in tile-major HBM matrix layout so each 64 x 64 weight tile is physically contiguous and the generated stride matches that layout.
• Add compact expert-major VRAM gather and fixed-route grouped expert lowering. Tokens routed to the same known expert share one gate/up/down weight load instead of reloading the same weights for every routed pair.
• Add an opt-in two-panel ping-pong lowering path using existing H_PREFETCH_M and matrix instructions, with strict Matrix-SRAM capacity checks.
• Restore stage attribution to expert_weight_prefetch after dynamic expert-address calculation.
• Add CI coverage for storage layout, compact grouping, one-load-per-expert behavior, stage attribution, invalid grouping, and ping-pong capacity/order.

Scope and compatibility

This is a compiler-only, opt-in change. It does not add or rename opcodes, change RTL, or modify the default row-major/per-pair/blocking path. A direct comparison against main produced the same 74 non-comment instructions and the same SHA256 for the default path.

Grouped lowering is intentionally limited to fixed-route or trace-replay programs where expert groups are known when code is generated. It does not claim runtime grouping for device-selected routes; that still needs a runtime sort/segment mechanism.

The rejected experimental deep-ring and cross-expert panel-pool implementations are not included.

Validation

• Python 3.12 minimal CI environment: 10 passed in test_moe_expert_grouping.py.
• MoE compiler guards: 36 passed.
• Full aten/tests: 94 passed; 3 failures reproduce unchanged on origin/main (Transformers router API drift, an existing packed-router comment assertion, and the existing slow quantization threshold).
• git diff --check: clean.
• No developer-local absolute paths or experiment artifacts are included.