fix: gracefully truncate context on long conversations

[ArvindShah]

Summary

• Bug: generate() concatenates all context segments unconditionally, then raises a ValueError only after tokenization is complete. For multi-turn conversations this causes unpredictable crashes once the accumulated frames exceed the 2048 limit.
• Fix: Implements a sliding-window truncation strategy that drops the oldest segments to fit within the frame budget, keeping the most recent conversational context intact.
• Warning: Emits a warnings.warn() when segments are dropped, so callers can detect and handle the truncation without catching exceptions.

Changes

• Compute the frame budget upfront (max_seq_len − max_generation_len − gen_text_frames)
• Tokenize context segments individually, tracking per-segment frame counts
• Drop oldest segments iteratively until total context fits within budget
• Improved error message when input text alone exceeds the budget
• No API changes — existing callers work without modification

Test plan

• Run python run_misotts.py with the default 4-turn conversation (fits within context — no truncation, same output as before)
• Extend the conversation list to 10+ long utterances to trigger truncation — verify a UserWarning is emitted and audio still generates correctly
• Pass a very long single text string that exceeds the budget on its own — verify a clear ValueError is raised with an actionable message