[Feature] Add GLM-5.3-Flash F3: Kimi Delta Attention (KDA)

[jayhenry]

Stack #2179 (bottom to top):

1. [Fix] Preserve DeepEP layout events in synchronous dispatch #2177 fix/deepep-layout-event → main
2. [Feature] Add GLM-5.3-Flash F0: 25B cropped reference checkpoint builder #2105 feat/glm53flash-materialize-full-f0 → fix/deepep-layout-event
3. [Feature] Add GLM-5.3-Flash F3: Kimi Delta Attention (KDA) #2106 feat/glm53flash-f3-kda → feat/glm53flash-materialize-full-f0 ← you are here
4. [Feature] Add GLM-5.3-Flash F4: mHC four-stream residual #2107 feat/glm53flash-f4-mhc → feat/glm53flash-f3-kda
5. [Feature] Add GLM-5.3-Flash F5: NoPE DSA + KPool indexer + clamped SwiGLU #2108 feat/glm53flash-f5-nope-dsa → feat/glm53flash-f4-mhc
6. [Feature] Add GLM-5.3-Flash F1: VL data preprocessing pipeline #2109 feat/glm53flash-f1-vl-data → feat/glm53flash-f5-nope-dsa
7. [Feature] Add GLM-5.3-Flash F2: vision tower + projector (eager) #2110 feat/glm53flash-f2-vision-tower → feat/glm53flash-f1-vl-data
8. [Fix] Correct generic MTP future embeddings and empty-shard loss #2178 fix/mtp-local-future-embeddings → feat/glm53flash-f2-vision-tower
9. [Feature] Add GLM-5.3-Flash F6 core: text model + MTP + compose model #2111 feat/glm53flash-f6-text-moe → fix/mtp-local-future-embeddings

Base is #2105's branch (layer 2). Review only this PR's own diff.

---

Summary

Stack layer 3/9 of GLM-5.3-Flash support (base: layer 2, F0 + materialize_full).

The KDA compile, precision, kernel-gradient and sequence-parallel fixes previously in F6 are now included in this layer. The shared fp32_linear utility is here because the KDA gate uses it.

Adds xtuner/v1/module/attention/kda.py (KDAConfig / KimiDeltaAttention) with module structure and parameter names matching the published checkpoint layout (separate q/k/v_conv1d, low-rank f_a_proj->f_b_proj forget gate with fp32 A_log/dt_bias, low-rank g_a_proj->g_b_proj output gate).

Root cause of a real correctness bug found while implementing: the installed fla 0.4.2's chunk_kda/fused_recurrent_kda call convention differs from the xtuner-ncp-k3 reference — it doesn't accept A_log/dt_bias/use_beta_sigmoid_in_kernel, and silently swallows them into **kwargs without raising, producing a KDA that trains but computes the wrong gate. Fixed by computing the gate externally via fla.ops.kda.gate.fused_kda_gate before calling the kernel, matching HF's own Glm5NextTextLinearAttention reference forward. A regression test pins these three kwargs' absence from chunk_kda's signature so this can't silently regress.

Also relaxes TransformerConfig.linear_attention to GatedDeltaNetConfig | KDAConfig | None (mypy-strict plumbing only, no behavior change for existing GatedDeltaNet layers).

The Read the Docs configs now mock einops, which KDA imports while Sphinx loads XTuner with docs-only dependencies. English and Chinese HTML builds passed locally with Python 3.12 and Sphinx 6.2.1.

Test Plan

tests/model/test_glm53_kda.py (5 tests, single/2-GPU): gate parity against fla's naive reference, module-level parity against HF's Glm5NextTextLinearAttention (weights bridged per the published<->HF key mapping), packed multi-document isolation, and Ulysses SP parity.

[jayhenry]

是否可以和 Qwen3.5 的合并？

[jayhenry]

这样做和使用 mixed_qkv相比，哪个好？

[jayhenry]

直接让 self.q_conv1d 等支持 SP 是否更好？