feat: support Intern-S2-Preview TS encoder

[CUHKSZzxy]

Summary

• Add InternS2 Preview encoder-only time-series support in a separated PyTorch/VL model path.
• Keep the InternS1-Pro time-series encoder rename and shared Whisper K-proj bias loader fix.
• Skip packaged time-series forecaster checkpoint tensors during loading without exposing forecast API/runtime paths.

Validation

• python -m py_compile on touched Python files
• pytest tests/test_lmdeploy/test_vl/test_preprocess_utils.py -q
• pre-commit hooks from git commit

Assistance

Assisted with Codex + GPT-5.6-Sol xHigh, reviewed manually

[Copilot]

Pull request overview

This PR adds dedicated Intern-S2-Preview time-series (TS) support by splitting it into a separate VL/PyTorch model path, while also improving TS metadata propagation and fixing Whisper packed-QKV loading for TS encoders.

Changes:

• Introduce new InternS2Preview VL + PyTorch model implementations, including a dedicated TS encoder and forecaster-weight skipping during load.
• Propagate ts_channels through preprocessing/expanded multimodal items to support multi-channel TS batching.
• Remove TS handling from generic Qwen3.5 VL/PyTorch paths and update docs + tests to reflect Intern-S2-Preview TS support.

Reviewed changes

Copilot reviewed 17 out of 18 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/test_lmdeploy/test_vl/test_preprocess_utils.py	Adds coverage for TS expansion metadata (ts_channels, ts_token_id).
lmdeploy/vl/model/qwen3_5.py	Removes InternS2Preview and TS preprocessing from the generic Qwen3.5 VL model path.
lmdeploy/vl/model/preprocess_utils.py	Preserves ts_channels when expanding TS multimodal items.
lmdeploy/vl/model/interns2_preview.py	Adds a dedicated VL model wrapper for InternS2Preview with TS preprocessing/token counting logic.
lmdeploy/vl/model/builder.py	Registers the new InternS2Preview vision model for construction.
lmdeploy/vl/model/base.py	Adds ts_channels modality mapping and constrains requests to one TS input.
lmdeploy/pytorch/models/whisper.py	Adds helper to synthesize missing Whisper K-proj bias for packed QKV loading in TS modules.
lmdeploy/pytorch/models/qwen3_vl.py	Removes TS MultiModalData construction from generic Qwen3VL input processing.
lmdeploy/pytorch/models/qwen3_5.py	Removes TS forward/generation plumbing from the generic Qwen3.5 PyTorch model path.
lmdeploy/pytorch/models/qwen3_5_moe.py	Stops building legacy TS module in Qwen3.5-MoE and skips forecaster tensors only.
lmdeploy/pytorch/models/module_map.py	Maps InternS2Preview architectures to the new PyTorch model implementation.
lmdeploy/pytorch/models/interns2_preview.py	Adds InternS2Preview PyTorch model + input processor with TS batching/padding support.
lmdeploy/pytorch/models/interns2_preview_ts_encoder.py	Implements InternS2Preview TS encoder stack (Whisper/BERT-derived components).
lmdeploy/pytorch/models/interns1_pro.py	Renames TS encoder import and applies the Whisper K-proj bias synthesis during weight load.
lmdeploy/pytorch/models/interns1_pro_ts_encoder.py	New/renamed TS encoder implementation for InternS1-Pro.
lmdeploy/pytorch/models/bert.py	Adds a lightweight BERT implementation used by the new TS encoder stack.
docs/zh_cn/multi_modal/multimodal_inputs.md	Updates Chinese docs to include Intern-S2-Preview TS support and clarify sampling rate.
docs/en/multi_modal/multimodal_inputs.md	Updates English docs similarly.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.