fix(glm): inline GLM-MoE-DSA routing to survive transformers without route_tokens_to_experts (#144)

[drunkcoding]

Problem (#144)

SyncGlmMoeDsaMoEBlock reused HuggingFace's internal GlmMoeDsaMoE.route_tokens_to_experts (bound to the offload block's self). Investigation showed the GLM routing architecture changed across the pinned transformers>=5.3,<6 range:

• transformers 5.12: GlmMoeDsaTopkRouter.forward returns raw router_logits; GlmMoeDsaMoE.route_tokens_to_experts(logits) does the grouped-sigmoid-top-k and returns (indices, weights).
• transformers 5.15 (what CI resolves): route_tokens_to_experts is removed; the routing moved into GlmMoeDsaTopkRouter.forward, which now returns a 3-tuple (router_logits, topk_weights, topk_indices).

So GLM was genuinely broken on 5.15 (not just a test issue): #147's guard turned the crash into a RuntimeError and the test module-skipped, hiding it. The self.gate(...) output being a tuple on 5.15 is exactly why the block's routing failed ('tuple' object has no attribute 'sigmoid').

Fix — version-robust routing

_route is now adaptive:

• 5.15+: self.gate(...) returns (router_logits, topk_weights, topk_indices) → unpack and use directly.
• 5.12: self.gate(...) returns raw logits → apply an inlined _route_tokens_to_experts (a faithful copy of 5.12's grouped-sigmoid-top-k).

The fragile GlmMoeDsaMoE reference (and import) is removed, so routing no longer depends on an unstable HF internal. test_glm_routing.py: the module-level skip becomes a per-test skipif on test_routing_parity only (that test compares against the HF method, absent on 5.15), so the other GLM tests now run and exercise the block on all builds.

Verification

• transformers 5.12 (local, real): pytest test_glm_routing.py → 5 passed. Direct numerical parity: inline routing == HF route_tokens_to_experts (torch.equal indices, allclose atol 1e-6 weights). Method-absent simulation: construct + forward OK.
• transformers 5.15 (downloaded the wheel, read the exact source): verified the block's adaptive _route unpacks the real (router_logits, topk_weights, topk_indices) contract and forward() produces correct-shape output.
• ruff format/check, py_compile, unit-test collection: clean.

Note: a pre-existing environment-only failure (test_glm_moe_dsa.py::test_glm_cpp_dequant_matches_python, missing compiled _store.dequant_fp8_blockwise) is unrelated (fails identically on clean dev).

Closes #144.