feat: LLM-triggered switch_language tool for multilingual conversations

[san0808]

Summary

• Adds switch_language as a built-in LLM tool (same pattern as transfer_call) that lets the LLM switch ASR/TTS languages mid-conversation via function calling
• Adds TranscriberPool and SynthesizerPool — thin wrappers that hold multiple transcriber/synthesizer instances and expose a switch(label) method for instant language switching
• Multilingual setup paths in __setup_transcriber / __setup_synthesizer that read a multilingual config block and build pools
• Tool is auto-injected into api_tools when pools are active (with enum constraint listing available languages) — no manual config needed
• Pre-call filler message is suppressed for switch_language so the switch is seamless
• On tool call: pools switch → tool result appended to history → LLM re-invoked so it continues in the new language

Flow

LLM emits switch_language(language="hi")
  → TaskManager.__execute_function_call intercepts it
    → calls self.switch_language("hi") → pools switch active member
    → appends tool result to conversation history
    → re-invokes LLM (which now responds in Hindi)

Files changed

• bolna/constants.py — SWITCH_LANGUAGE_TOOL_DEFINITION schema
• bolna/agent_manager/task_manager.py — __inject_switch_language_tool(), handler in __execute_function_call, switch_language() method, multilingual pool setup paths
• bolna/helpers/utils.py — suppress pre-call filler for switch_language
• bolna/transcriber/transcriber_pool.py — new TranscriberPool class
• bolna/synthesizer/synthesizer_pool.py — new SynthesizerPool class
• bolna/transcriber/__init__.py, bolna/synthesizer/__init__.py — pool exports