Fix sensory buffer loop for oversized messages

[taixiansen2]

Summary

Fixes an infinite loop in SenMemBufferManager.add_messages() when a single user message is larger than the sensory memory buffer's max_tokens while the current buffer is empty.

In the previous logic, this edge case called cut_with_segmenter() on an empty buffer. No segment was emitted and the oversized message was not added to processed_messages, so it was never removed from big_buffer. The outer while self.big_buffer loop then processed the same message forever.

This PR handles that edge case by emitting the oversized user turn as a singleton segment, optionally including the following assistant message, and consuming those messages from big_buffer.

Changes

• Return early for empty messages and empty internal buffers.
• Consume oversized single user messages when the current buffer cannot be segmented.
• Preserve the following assistant message with the oversized user message when available.
• Guard turn construction against incomplete or non-user/assistant pairs.
• Fix force_segment=True cleanup to clear the whole remaining buffer with len(self.buffer) instead of len(boundaries).
• Add a small regression test for the oversized-message case and force-segment cleanup.

Validation

• Parsed the changed Python files with ast.parse.
• Ran the new regression test functions directly with PYTHONPATH=src because pytest is not installed in the local Windows Python environment:

PYTHONPATH=src python - <<'PY'
from tests.test_sensory_memory import (
    test_oversized_single_user_message_is_consumed,
    test_force_segment_flushes_remaining_buffer,
)

test_oversized_single_user_message_is_consumed()
test_force_segment_flushes_remaining_buffer()
PY

In our downstream LongMemEval reproduction, this fixes a run that previously stalled at Ingest 01493427: 105/242; after the fix, the case completed successfully.