Fix animation command tags leaking into TTS output

[bedge117]

GetActionsFromString did not guard against malformed command tags, causing raw command stems (e.g. "playAnimationWI") to be spoken aloud instead of parsed as actions. When an LLM emitted a tag missing the "||" separator (e.g. "{{playAnimationWI}}"), the parser indexed cmdPlusParam[1] without a length check, mishandling the tag and leaving the command stem glued to the spoken text.

This is common with local/open-source models (Ollama, etc.) that don't format the {{command||parameter}} structure as reliably as OpenAI.

Changes:

• Guard cmdPlusParam access with a length check so tags missing "||" no longer mis-parse.
• Add stripCommandStems() to scrub leaked command keywords (playAnimation, playAnimationWI, getImage, newVoiceRequest) and stray brace debris from spoken text before it reaches TTS.
• Valid, well-formed tags still fire their animations as before; only malformed/leftover tag text is removed from speech.