[build] Sync sglang-kt submodule (v0.6.4)

[github-actions]

Automated sync of third_party/sglang submodule to latest main.

Old ref: bc7f0058fafe6c05dc6ad52d6b1c854af2a673b2
New ref: ad29d4d2971532a2e2ab47398b6de032ad007b6c
sglang-kt version: 0.6.4

Commits included

ad29d4d29 fix(kv-cache): guard compressed pool initialization for non-SWA models (#95)
4afb5b541 fix: initialize hybrid KV flags for non-SWA models
94d719387 fix(kimi-k25): avoid the unused cuDNN Conv3d startup guard
4b58c35ff test(packaging): handle encoded Python source in import audit
ba7f7a9f1 fix(packaging): vendor compressed-tensors for KT-only installations
df416c647 release: require Accelerate carrier for the complete post4 stack
2139e8d42 release: align dependencies with the post4 main stack
112baf622 fix(kt): support Kimi RAWINT4 LoRA adapter serving (#94)
f2341d932 fix(pypi): converge verified GLM post3 runtime into main (#88)
fa7c29d66 fix(dsv4): honor --max-total-tokens for the compressed-attention pool (#83)
0980b6da2 fix(openai): support DSV4 reasoning in Responses API (#92)
5d861db9a perf(dsv4): parallelize SM86 non-expert decode (#87)
3424f35d0 fix(glm5-next): accept MTP index sharing metadata (#85)
d26654844 Fix double routed scaling in KT EP MXFP4 experts (#86)
6df1c9cc8 fix(lora): support all-target KT MoE adapters (#82)
5d16c520f feat(glm5-next): support GLM-5.3-flash serving (#81)

---

This PR was created automatically by the sync-sglang-submodule workflow.