add DFlash2 npu training recipe

[curnane-lab]

Motivation

Follow-up to #772: make DFlash 2 trainable on Ascend NPU by adding the missing recipe layer — one draft config, one external-server recipe, and a tutorial section. DFlash 2 reuses the DFlash capture path and the NPU disaggregated plumbing from #722 unchanged. Mirrors the Domino/DSpark NPU recipes in #726 / #727.

Modifications

• configs/qwen3.5-4b-dflash2.json: DFlash2DraftModel config for Qwen3.5-4B — 5 full-attention layers on the same target layers as the 4B DFlash config ([1, 8, 15, 22, 29]), 16, conv_kernel_size: 2, conv_group_size: 16, selector_rank: 256, selector_top_k: 16`.
• examples/configs/online/disaggregated/external/qwen3.5-4b-dflash2-online-npu.yaml: field-aligned with the existing DFlash NPU recipe, except attention_backend: sdpa (flex_attention is unavailable on Ascend) and dflash2_selector_loss_alpha: 1.0.
• docs/basic_usage/Ascend/ascend_npu.md: a "Training DFlash 2 drafts" section — capture setup unchanged from DFlash; sdpa mandatory; the convolutions and selector are stock torch ops; serving needs SGLang with DFlash 2 ([Spec] DFlash2: local convolution + candidate selector sglang#35371, merged).

Related Issues

Follow-up to #772; builds on #722; serving side is sgl-project/sglang#35371.

Accuracy Test

Config / recipe / docs only; no model-side code. End-to-end validation on Ascend NPU with the checked-in recipe (Qwen3.5-4B target + DFlash 2 draft, ~1.75k optimizer steps): train/loss ~6.9 → ~1.25, train/acc ~0.09 → ~0.71, selector_loss 2.05 → ~0.25, selector_accuracy ~0.36 → ~0.93, selector_coverage ~0.44 → ~0.90; bounded grad_norm, no NaN or stall.

Benchmark & Profiling

N/A (no performance-affecting code changes).

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests. (Recipe-only change; no new code paths to cover.)
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results. (Training-side NPU validation in Accuracy Test; serving benchmark out of scope.)
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.