Summary

Implements concurrent speaker support for SelectorGroupChat as outlined by @ekzhu in #5395:

• selector_func can now return str | list[str] | None — returning a list selects multiple speakers to respond concurrently
• New max_concurrent_speakers parameter (default 1) enables model-based multi-speaker selection
• When multiple speakers are selected, all receive RequestToSpeak and must respond before the next selection round
• allow_repeated_speaker=False now correctly excludes all previous concurrent speakers from the next round

Backwards compatible: default behavior is unchanged (max_concurrent_speakers=1, selector_func returning str | None).

Closes #5395

Changes

• Updated SelectorFuncType signatures to accept str | list[str] | None return type
• Added max_concurrent_speakers parameter to SelectorGroupChat and SelectorGroupChatManager
• Added _select_multiple_speakers() method for model-based concurrent selection
• Changed _previous_speaker: str | None → _previous_speakers: list[str] for correct exclusion tracking
• Updated SelectorManagerState with backwards-compatible previous_speakers field
• Updated serialization (save_state/load_state) for backwards compatibility with existing state
• Updated SelectorGroupChatConfig and _to_config/_from_config for declarative support

Test plan

• All 30 existing selector tests pass (no regressions)
• New test: test_selector_group_chat_concurrent_speakers_with_selector_func — selector_func returns ["agent1", "agent2"]
• New test: test_selector_group_chat_concurrent_speakers_with_model — model returns "agent1, agent2" with max_concurrent_speakers=2
• pyright: 0 errors, 0 warnings
• Full test suite: 91/91 pass