feat: add double-buffer context management

[marklubin]

Summary

Adds opt-in double-buffer context management that produces higher-quality chat history summaries by checkpointing early (at 60% of max_chat_history_tokens) while the model still has full attention, rather than waiting until the history is too big and attention is degraded.

How it works:

1. At 60% capacity, a background thread runs ChatSummary.summarize() on a snapshot of done_messages
2. Normal operation continues uninterrupted (the checkpoint is fire-and-forget)
3. At 85% capacity, instead of the standard stop-the-world summarization, the pre-computed summary is swapped in

The feature is opt-in via --double-buffer and the existing summarization flow is completely unchanged when not enabled.

New files:

• aider/double_buffer.py: DoubleBufferManager with background checkpoint via threading, Phase enum
• tests/basic/test_double_buffer.py: 18 tests (phase transitions, thresholds, error handling)

Modified files:

• aider/coders/base_coder.py: summarize_start() delegates to _double_buffer_check() when --double-buffer is enabled
• aider/args.py: --double-buffer, --checkpoint-threshold (default 0.60), --swap-threshold (default 0.85)

Design:

• Reuses existing ChatSummary for summarization (no new LLM calls)
• Background thread mirrors existing summarizer_thread pattern in base_coder.py
• Falls back to standard summarization if no double-buffer action applies

Algorithm description: https://marklubin.me/posts/hopping-context-windows/

Test plan

• 18 new unit tests pass (pytest tests/basic/test_double_buffer.py -v)
• No regressions in existing tests
• black --line-length 100 --preview clean
• isort --profile black clean
• flake8 clean (no warnings)
• Manual test: run with --double-buffer --verbose on a long session and observe checkpoint/swap log messages

[CLAassistant]

All committers have signed the CLA.