[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.

Merges the latest test-nemotron layout refactor and places the implementation, reference, tests and benchmark under the canonical directories. The operator and reference are byte-identical to the previous PR version; test and benchmark computations are unchanged. Registration lives in rl_engine.runtime.registry, preserving the upstream compatibility entry point.

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 remain proposed for section 4.4 review. Expert MLP/combine and distributed communication are outside this operator.

Local migration validation:

• Focused router/reference/registry/layout tests: 41 passed, 81 CUDA skipped; upstream layout suite: 32 passed, 1 GPU extension skipped. These sets overlap.
• Additional dispatch, stateless-executor, cross-configuration and tolerance tests: 87 passed.
• Strict documentation build, scoped formatting/lint/types, wheel build and imports outside the source checkout pass.
• Full-package MyPy reports the same 14 errors as the target branch in the same local environment, with no new diagnostics.

Under the canonical layout, all 106 single-H20 router/reference/registry tests passed without skips. A fresh cuBLAS + TE 2.19 index/eager replay completed 256 route/full measurements with accuracy qualification; its 128 full-operator rows retain paired samples and regressions. BF16 random forward+backward at 8192/32768 tokens measured 1.087–1.092x; FP32 remained near parity and small-token regressions persist. Raw measurements are retained separately for review. The unchanged operator also has historical 36-case bitwise comparisons and four-tool Compute Sanitizer qualification; those checks were not rerun during this layout replay.

Target-branch CI and maintainer approval of the proposed contract remain pending.