[Nemotron][CUDA] Add deterministic moe_router_dispatch - #504

Implements the SM90 CUDA moe_router_dispatch work item of #434: FP32 projection, sigmoid, bias-corrected top-six selection, normalization and deterministic token packing, with backward for routing weights and packed payloads. Supports BF16/FP32 inputs with FP32 router weights at Nano's 2688-hidden, 128-expert geometry.

Adds a numerical reference, CPU/GPU correctness and batch-invariance tests, a cuBLAS + Transformer Engine benchmark, documentation and CI wiring for test-nemotron. The ordering and fixed arithmetic policy are proposed for section 4.4 review. Expert MLP/combine and distributed communication are outside this change.

Validation:

• Single-H20 qualification of the unchanged operator: 106 tests passed; output and gradient bits matched the prior schedule in 36 cases; four Compute Sanitizer tools passed.
• After consolidating the GPU test files, local checks: 25 passed, 81 CUDA skipped. Original test function bodies and parameterization are preserved; no new GPU run was performed for this consolidation. Formatting and lint checks pass.
• Random BF16 forward+backward at 8192/32768 tokens is 1.08-1.09x faster than cuBLAS + TE 2.19 index dispatch/autograd. FP32 is near parity; small-token regressions are reported. Both providers use eager timing. The benchmark records raw timing samples, incremental peak memory and source hashes; the measured matrix is retained separately for review.

The target branch still has the eight documentation link warnings addressed by #458 on main. This change adds no documentation warnings; strict docs passes locally with a separate fix for those links. Target-branch CI and maintainer approval of the proposed contract remain pending.