feat(deepseek-v4): CP adaption and gfx942 (MI308X/CDNA3) adaptation — 128k CP + full-model multi-node SFT

[aaab8b]

What

Make DeepSeek-V4-Flash trainable on gfx942 / CDNA3 (MI308X). The upstream
sparse-MLA kernels target gfx950/CDNA4; this PR adapts them plus the long-context
and multi-node paths so the full model trains on CDNA3.

gfx942 kernel fixes

• DSA sparse-MLA backward LDS overflow (compile-time): the 512-latent Q tile
  plus Triton's stage double-buffering asks for 73728 B > the 64 KB CDNA3 LDS
  budget. Fixed with num_stages=1 (PRIMUS_DSA_BWD_NUM_STAGES), an
  amd_pingpong_disabled() compile scope that disables primus_turbo's global
  ping-pong / async-copy knobs (which double-buffer LDS and are a measured
  pessimization for these kernels), and tunable tile sizes as a fallback.
• Forward pinned to the non-ping-pong schedule; backward skips provably-zero
  rope MMAs.

Long-context / context parallelism

• CSA/HCA CP: left-boundary KV exchange (P2P) + global compressed-pool
  all_gather (deepseek_v4_cp.py).
• Indexer scoring gains Q_OFFSET so visibility is judged on the query's
  global position under CP (0 without CP → identical to before).
• Streaming chunked top-K (PRIMUS_INDEXER_TOPK_CHUNK) avoids materializing the
  full [B,S,P] score row; full-Triton indexer path avoids an eager einsum that
  would materialize [B,S,H,P].

Multi-node full-model SFT

• NNODES=2/3 branches with PP layouts; optimizer CPU offload with a tunable
  fraction; socket-network note: NCCL_ALGO=Ring fixes a Tree-over-TCP
  cross-node all_reduce deadlock on fabrics without RDMA.

Scripts & docs

One-click launchers examples/deepseek-v4/gfx942/run_128k_dense_hca_csa.sh and
run_full_4k_multinode.sh, plus READMEs.

Verification

• 128k, 4-layer (CP=8/TP=1): 42.30 GB peak.
• Full model, 4k, 3 nodes: 10-step and 100-step runs, loss 11.90 → 9.19
  monotone, grad norm 20 → 2.8, 0 nan iterations, exit 0.

All changes are gfx942-gated or env-var-defaulted-off and do not change
gfx950 numerics or behavior.

Test plan

• pre-commit run --all-files green on the changed files
• 128k 4-layer smoke on 1×8 MI308X
• full-model 4k 10-step & 100-step on 3×8 MI308X (no NaN)