DSH context compression/summarization plugin

Key features

• Estimate the current session token usage through ctx.tokenMeter
• Every compression will perform before/after verification: totalTokens after compression must be strictly smaller than before compression
• Global checkpoint, not just one section: each trigger will make "all earlier history" (from the earliest message to before retaining the tail) into a global checkpoint at once
• When the provider clearly reports that the context has exceeded the limit, it will first summarize the compression in detail, and then automatically retry the current round of requests
• When reaching 80% of the model window (thresholdRatio defaults to 0.8), a detailed summary is automatically made before the next step: compress all the earlier history into a detailed Chinese checkpoint using LLM, replace the old messages, and only keep the original tail of the most recent retainRatio (default 16%)