perf: use BTreeMap::range() and incremental memory tracking in reassembly

[Zious11]

Summary

Performance improvements for the TCP reassembly engine identified during PR #10 review.

Items

1. Use BTreeMap::range() for overlap detection instead of iterating all segments (segment.rs:85). Only examine segments that could overlap with the new segment's range.
2. Maintain buffered_bytes incrementally in FlowDirection instead of O(n) summation per insert (segment.rs:62).
3. Maintain total_memory incrementally instead of update_memory() O(n) recomputation across all flows (mod.rs:379).

Impact

Current approach is O(n) per packet per flow for memory tracking and O(n) for overlap scan. With max_segments=10K, this is bounded but wasteful. Incremental tracking makes all three O(1).

Acceptance Criteria

• Overlap detection uses range(candidate_start..new_end)
• FlowDirection tracks buffered_bytes as a running counter
• total_memory updated incrementally on insert/flush/remove
• No performance regression in existing tests