[CI] Avoid Hugging Face Hub API calls when the local model cache is complete

[hnyls2002]

• _validate_sharded_model: derive the index name from shards named model.safetensors-0000N-of-0000M.safetensors (Qwen3.5) as model.safetensors.index.json; CI cache validation of these models always failed and fell through to Hub API calls.
• find_local_repo_dir and the custom cache_dir lookup: resolve a branch or tag revision (e.g. the draft model's default "main") through refs/<revision>, so draft models hit the local snapshot instead of going online on every load.
• download_weights_from_hf: when listing the repo on the Hub raises HfHubHTTPError (e.g. 429), pick the weight format from the local snapshot instead of failing startup.
• CI tests: in-process Engine startups (CustomTestCase, CI only) run with HF_HUB_OFFLINE=1 when the model, tokenizer, and draft repos all validate in the local cache, sharing the per-model validation with popen_launch_server. LoRA startups stay online, as in popen_launch_server.
• CI install: drop the per-job kernels download / kernels lock step; the Hub FA3 kernel is only used with SGLANG_USE_SGL_FA3_KERNEL=0.
• Add a regression test for the shard index name in test/manual/test_weight_validation.py.

No behavior change outside CI, except that an explicit branch or tag revision now loads from the local cache when refs/<revision> is present.

---

CI States

Latest PR Test (Base): 🚫 Run #37856132853
Latest PR Test (Extra): ❌ Run #37856132358
Latest PR Test (AMD ROCm 10): 🚫 Run #37856132473

[hnyls2002]

/tag-and-rerun-ci

[hnyls2002]

/rerun-test test_lora_qwen3_5_35b_a3b_logprob_diff.py test_eagle_dp_attention.py test_deepep_small.py test_vlm_tp4.py test_score_engine.py test_embedding_models.py test_qwen35_deterministic.py test_lora_qwen3_5_4b_logprob_diff.py

[github-actions]

🚀 4-gpu-h100 (5 tests): ❌ View workflow run

cd test/ && python3 registered/lora/test_lora_qwen3_5_35b_a3b_logprob_diff.py
cd test/ && python3 registered/spec/eagle/test_eagle_dp_attention.py
cd test/ && python3 registered/ep/test_deepep_small.py
cd test/ && python3 registered/vlm/test_vlm_tp4.py
cd test/ && python3 registered/attention/test_qwen35_deterministic.py

🚀 1-gpu-5090 (2 tests): ✅ View workflow run

cd test/ && python3 registered/prefill_only/test_score_engine.py
cd test/ && python3 registered/prefill_only/test_embedding_models.py

🚀 1-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/lora/test_lora_qwen3_5_4b_logprob_diff.py