Implement adaptive chunked compaction

[bug-ops]

Parent: #740 (P1)

Problem

Single summarization request for large context may exceed model token limits or produce lower quality summaries.

Solution

• Split compaction input into token-budgeted chunks (max ~4K tokens each)
• Summarize chunks in parallel via tokio::join!
• Merge partial summaries with a final consolidation prompt
• Add 1.2x safety margin to token budget calculations
• Detect oversized single messages (>50% of chunk budget) and handle separately

Affected crates

• zeph-core (compaction logic in agent/context.rs)

Acceptance criteria

• Compaction works correctly with >32K token contexts
• Parallel chunk summarization
• Final merge produces coherent summary
• Graceful handling of oversized messages