feat(dflash): add fused CE and total anchor budget

[Slyne]

What

• add a local-microbatch max_total_anchors budget so anchor construction has a predictable memory ceiling
• add chunked fused LM-head + cross-entropy execution that avoids retaining full-vocabulary logits
• preserve dense/fused metric and gradient parity, including TP and ancestor-owned FSDP2 target heads
• normalize weighted DFlash-family objectives over the draft data-parallel group, with separate DFlash2 base and selector denominators
• thread the anchor budget through DFlash2, Domino, and JetSpec and document the new controls

The shipped Qwen3 DFlash examples enable fused CE to reduce peak training memory. The recipe API still defaults it to disabled. Fused CE supports loss_type: dflash; variable-prefix training must disable it.

Testing

• 243 passed across the focused DFlash, DFlash2, Domino, JetSpec, loss, draft, TP, FSDP, and recipe suites
• real two-rank DDP/no_sync gradient parity with uneven per-rank base and selector counts, including a rank with no selector examples
• Ruff formatting/lint and git diff --check

Review

• independent review-agent loop: no actionable findings
• Claude Code Opus, maximum effort, repeated after fixes: no actionable findings

[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.

[github-actions]

🌿 Preview your docs: https://nvidia-preview-preview-ee03a5364fff.docs.buildwithfern.com/nemo/automodel