feat(speculative): context parallelism for draft-model training

[kashif]

Adds context parallelism (cp_size>1) to the DFlash/Domino/JetSpec and DSpark draft trainers for long-context training, and fixes the same path in EAGLE-3.

The approach follows EAGLE-3's existing seam: only the frozen target's forward is sharded across the cp mesh, and its captured hidden states are gathered back to the full sequence before the draft runs — so the draft's block/anchor masks are never sequence-split. The draft shards over the dp submesh; cp ranks in a dp group share data and draft weights.

Two things that bit me and are worth a careful look in review:

• The drafts sample anchor positions from the global RNG, so cp (and tp) replicas would diverge if seeded per global rank. Now seeded per dp rank so they stay in sync.
• torch's context_parallel ring doesn't actually fire for a plain HF target — q/k/v reach SDPA as local tensors, so each rank silently attended only to its own sequence shard. Since the target is frozen (forward-only), I all-gather K/V and attend the local Q against the full sequence with a global causal mask (new attach_cp_kv_gather_hooks). This needs the HF torch-SDPA path, so cp_size>1 requires target_force_hf=true.

Validated on 2 GPUs (Qwen3-0.6B): with the fix, per-step draft loss matches the 1-GPU baseline within bf16 noise (mean signed +0.18%, symmetric above/below), versus a systematic +2–7% before. Unit tests pass.

Defaults (cp_size=1) are unchanged, and DSpark CP is gated to the dense Qwen3 target (the MoE targets already run under their own EP/FSDP mesh).

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[kashif]

One design question for reviewers on the attach_cp_kv_gather_hooks approach.

Right now the CP hook intercepts attention by swapping torch.nn.functional.scaled_dot_product_attention during each self_attn forward. That only works when the target lowers to torch SDPA, which is why cp_size>1 requires target_force_hf=true — a NeMo-custom target defaults to TE attention (and flex/tilelang/kernels-provided kernels would also bypass the swap), so the hook would silently no-op and each rank would fall back to local-only attention.

A more robust alternative: register a dedicated "cp_gather" entry in HF's ALL_ATTENTION_FUNCTIONS and select it via attn_implementation, instead of monkeypatching the torch symbol. That intercepts at the attention-interface level regardless of the underlying kernel, and would let us drop the force_hf requirement (and compose CP with faster attention backends). The trade-off is a bit more surface tied to the transformers attention-interface API across versions.

Worth doing in this PR, or fine as a follow-up? Leaning follow-up since the current gated version is correct where it's enabled, but happy to fold it in if preferred.

[khazic]

Taking a look at this now, will review shortly.

[khazic]

One PR-level item: none of the 13 commits carries a Signed-off-by trailer, and CONTRIBUTING requires DCO sign-off on every commit. A rebase with --signoff will fix it.

[khazic]

There are a few more issues that I am still reviewing in depth.