docs: add npu tutorial

[curnane-lab]

Motivation

The Ascend NPU managed-local stack landed in #722 but ships no documentation: the two SGLang patches, the device-visibility semantics, and the NPU-only constraints are currently only discoverable from PR descriptions. This PR adds the missing end-to-end Ascend tutorial, following the AMD ROCm template from #698, and ticks the docs checkbox left over from #722.

Modifications

• docs/basic_usage/Ascend/ascend_npu.md: new tutorial — installation (--no-deps + the base/companion capture patches applied in order), data preparation, online training against an external capture server, the one-command managed-local full stack (including a 6:10 capture/trainer override and rerun cleanup), multi-node role split, an NPU troubleshooting list (empty ASCEND_RT_VISIBLE_DEVICES rejected by the driver, USP/yunchang CUDA-only, wildcard Mooncake registration, ACL_ERROR_RT_CONTEXT_NULL, teardown drain TTL), and the validated 16-card A3 reference results from feat(npu): managed-local disaggregated DFlash online training on Ascend NPU #722 with training curves.
• docs/index.rst: register the tutorial in the Basic Usage toctree next to the ROCm one.

Related Issues

• Follow-up to feat(npu): managed-local disaggregated DFlash online training on Ascend NPU #722 (NPU managed-local stack; the reference numbers come from its validated run)
• Template follows [ROCm] AMD ROCm enablement — tutorial + managed_local Mooncake teardown/pin fixes #698 (ROCm tutorial)
• Recipes for Domino (feat(domino): add Qwen3.5-4B managed-local NPU disaggregated recipe #726) and MTP (feat: add mtp support #667) can be linked once they land

Accuracy Test

NA (documentation only). The tutorial's reference results reproduce the validated 16-card A3 run from #722 (loss ~6.0 -> ~2.0 over ~1.18k steps).

Benchmark & Profiling

NA (documentation).

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Update documentation /strings / example tutorials as needed, according to Writing Documentation.
• Add unit tests as outlined in the Running Unit Tests (NA, docs only).
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results (NA, docs only).
• 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 channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.