perf(engine): implement matsumoto line group partitioning

[HarshK97]

Problem

1. AST matching algorithms like TopDown and MatchUnmatchedLeaves compare nodes across the entire AST without line boundary constraints. On large files (1,000+ lines), this causes candidate combinatorial explosion where thousands of untouched subtrees are repeatedly compared against edited regions. So it takes over 200ms and 190MB of heap allocations per diff.
2. Unconstrained tree matching occasionally cross-matches identical syntax tokens (such as public void or return statements) in untouched lines against inserted or deleted code blocks elsewhere in the file.

What Changed

• Line Partitioning: Added LinePartition in internal/engine/partition.go, implementing the two-group line partitioning strategy from Matsumoto et al. ("Beyond GumTree"). It classifies lines into Group 1 (non-edited lines) and Group 2 (edited lines) using line-level LCS.
• Matcher Integration: Enforced CanMatch checks in TopDown and MatchUnmatchedLeaves so untouched subtrees only match corresponding untouched lines, while edited subtrees only match other edited regions. Cross-matching between untouched and edited lines is blocked.
• Unit Tests: Added unit tests in internal/engine/partition_test.go covering line classification, mapped line offset checks, and group boundary isolation.

Benchmark Results

• Large Diff Latency (Geometric Mean): Dropped from 115.30ms to 66.68ms (42.17% overall speedup, 1.73x faster)
• Large Diff Memory (Geometric Mean): Dropped from 113.10MB to 64.43MB (43.04% overall memory reduction, ~48.7MB saved per diff)