refactor: replace dict-based config with type-safe vendor classes

Summary

Replaces the old shorthand-string and dict-based vendor configuration (Agent(llm="openai/gpt-4")) with type-safe Pydantic vendor classes. This is a breaking API change — the previous AgentOptions TypedDict and all shorthand parsing functions are removed.

New API pattern:

from agoraio.wrapper import Agent
from agoraio.wrapper.vendors import OpenAI, ElevenLabsTTS, DeepgramSTT

agent = (
    Agent(instructions="You are a helpful assistant.")
    .with_llm(OpenAI(api_key="...", model="gpt-4"))
    .with_tts(ElevenLabsTTS(key="...", model_id="...", voice_id="...", sample_rate=24000))
    .with_stt(DeepgramSTT(api_key="..."))
)

What's new:

• vendors/ package with base classes (BaseLLM, BaseTTS, BaseSTT, BaseMLLM, BaseAvatar) and concrete implementations for ~30 vendors
• Pydantic BaseModel validation with extra = "forbid" on all vendor option classes
• Literal type aliases for vendor-specific sample rates (e.g., ElevenLabsSampleRate, CartesiaSampleRate)
• Avatar/TTS sample rate compatibility validation at both with_avatar() time and session start

What's removed:

• AgentOptions TypedDict
• _parse_llm_shorthand(), _parse_tts_shorthand(), _parse_stt_shorthand()
• avatar_types.py (validation logic moved into Agent and AgentSession)

Updates since last revision

• Removed stale exports from src/agoraio/__init__.py: AgentOptions, is_heygen_avatar, is_akool_avatar, validate_avatar_config, validate_tts_sample_rate were still referenced in the top-level __init__.py but no longer exist in the wrapper module. This was causing the mypy CI check to fail.

Review & Testing Checklist for Human

⚠️ This is a breaking change with no backward compatibility.

• Verify config dict keys match Agora API expectations: The vendor to_config() methods produce dicts with keys like "vendor", "params", "api_key", etc. These must exactly match what the generated Agora SDK types expect. Test with a real API call to ensure no field name mismatches — CI only checks types, not runtime correctness.
• Test avatar/TTS sample rate validation: Create an agent with .with_avatar(HeyGenAvatar(...)) (requires 24kHz) and .with_tts(ElevenLabsTTS(..., sample_rate=16000)) (wrong rate). Verify it raises ValueError at with_avatar() time. Also test the reverse order (TTS after avatar) to ensure AgentSession._validate_avatar_config() catches it.
• Verify Pydantic validation works: Try creating a vendor with invalid params (e.g., OpenAI(api_key="", model=123) or ElevenLabsTTS(key="", model_id="", voice_id="", sample_rate=99999)). Confirm Pydantic raises ValidationError with helpful messages.
• Check IDE autocomplete: Open the SDK in an IDE and verify that typing OpenAI( shows autocomplete for api_key, model, etc. with type hints.

Notes

• No unit tests were added in this PR. Consider adding tests for vendor class instantiation, to_config() output, and avatar/TTS validation logic.
• The OpenAITTS.sample_rate property always returns 24000 (hardcoded) since OpenAI TTS only supports 24kHz output.
• All vendor option classes use extra = "forbid" to catch typos in keyword arguments.
• Go SDK has equivalent changes in refactor: replace dict-based config with type-safe vendor classes agora-agents-go#6

Link to Devin run: https://app.devin.ai/sessions/756e000c856649af9c8655b0ef4e1038
Requested by: @fern-support