Collapse vendor configs into pydantic models for IDE autocomplete

[plutoless]

This PR has two related parts, both in service of IDE autocomplete for vendor configs.

Part 1 — Collapse vendor configs into pydantic models

Every STT/LLM/TTS/MLLM/avatar vendor config class (~63 across agentkit/vendors/) used def __init__(self, **kwargs: Any) and stashed a parallel *Options pydantic model, which erased all constructor type information — so VSCode/Pylance showed no autocomplete, no parameter types, and no hover docs when constructing e.g. DeepgramSTT(...).

This collapses each vendor class into its pydantic model: fields (with their Field(..., description=...), constraints, and validators) live directly on the class, and to_config() reads self.<field>. The *Options classes are deleted and the base classes become BaseModel + ABC. Result: full field autocomplete + hover docs + type-checking, with no new dependencies.

Behavior preservation

• to_config() output is byte-identical for every vendor (guarded by golden-master snapshots + the full existing suite).
• Two intentional changes:
  • FengmingSTT(**unexpected) now raises pydantic ValidationError instead of TypeError.
  • SenseTimeAvatar additionally accepts snake_case (app_id/scene_list) via populate_by_name=True; the wire output still emits appId/sceneList.

Notable design decisions

• OpenAI/Gemini-compatible LLMs (Groq, CustomLLM, VertexAILLM, and the CN Aliyun/Bytedance/DeepSeek/Tencent LLMs) are standalone copies rather than inheriting/delegating — removes the old __init__ kwargs mutation and gives each clean per-class hints (accepted duplication over coupling).
• TTS sample_rate: the base exposes resolved_sample_rate (reads getattr(self, "sample_rate", None)) to avoid a base member shadowing subclass sample_rate fields; agent.py's single consumer updated accordingly. Sarvam/Murf/CosyVoice keep their original accessor semantics via targeted overrides.
• HeyGenAvatar is a standalone copy of LiveAvatarAvatar with its deprecation warning moved to model_post_init.

Part 2 — Top-level type exports + agora_agent.cn

Part 1 made the classes typed, but the most natural import — from agora_agent import DeepgramSTT — still resolved to Any (the top-level package uses a lazy __getattr__ fallback). This part gives the global vendors real static types at the top level, and puts CN vendors in a dedicated module:

• Global vendors now autocomplete from the top level: from agora_agent import DeepgramSTT, OpenAI, MiniMaxTTS. Implemented by adding them to the top-level TYPE_CHECKING block only (they're already in agentkit.__all__, so runtime + __all__ are handled by the existing fallback/union — no _dynamic_imports/_ROOT_ALL churn).
• CN vendors get a new agora_agent.cn module with natural names: from agora_agent.cn import AliyunLLM, MiniMaxTTS, TencentSTT. This also resolves the three name collisions cleanly — agora_agent.MiniMaxTTS is the global class, agora_agent.cn.MiniMaxTTS is the CN class.
• Intentional change: SpatiusAvatar (a CN vendor previously exported top-level) is now cn-only — from agora_agent import SpatiusAvatar still runs (fallback → Any) but its typed path is from agora_agent.cn import SpatiusAvatar. XaiGrok/GenericAvatar were demoted to TYPE_CHECKING-only (behavior-neutral). One documented example in docs/guides/avatars.md updated; the broader CN-docs/alias-convention migration is a deliberate follow-up.

Test Plan

• poetry run pytest -rP . — 235 passed, 1 skipped
• poetry run mypy . — no new errors (only the 3 pre-existing core/ errors)
• Golden-master snapshot tests lock to_config() output for the high-risk classes
• Pyright (Cursor's engine): from agora_agent import DeepgramSTT and from agora_agent.cn import AliyunLLM resolve to real type[...] with 0 warnings; global vs CN MiniMaxTTS distinct
• Reviewer: spot-check autocomplete in VSCode/Cursor for one class per family

🤖 Generated with Claude Code