deps(treesitter): upgrade gotreesitter from v0.16.0 to v0.48.0

[HarshK97]

Problem

1. gotreesitter v0.16.0 had unbounded GLR parser stack splits on ambiguous syntax like Python blocks, causing severe parsing slowdowns up to ~400ms.
2. High allocation churn (13k to 19k allocations per diff) placed unnecessary memory pressure on the Go garbage collector.
3. CI benchmark execution was delayed because it waited on slow cross-platform test matrix runners (macOS and Windows) before starting.

What Changed

• Dependency Upgrade: Upgraded gotreesitter from v0.16.0 to v0.48.0.
• GLR Parser & Memory Optimization: Introduced GLR state-stack merging, iteration depth safety caps, O(N) leaf node matching, and scratch buffer pooling.
• Modular CI Jobs Architecture: Refactored .github/workflows/ci.yml into 4 dedicated, decoupled jobs:
  • lint: Fails fast in ~15s on syntax/formatting issues.
  • test-linux: Runs unit tests + code coverage on ubuntu-latest.
  • test-cross: Runs cross-platform unit tests on macos-latest & windows-latest in parallel.
  • benchmark: Depends on [lint, test-linux] (not test-cross), executing A/B benchmarks as soon as Linux unit tests pass with path filtering.
• CI Benchmark Formatting: Added a formatting script to parse benchstat output into a clean GitHub Step Summary report with geometric means, highlighted performance deltas, and collapsible breakdown tables.

Version Evolution

• v0.16.0: Base version. Suffered from unbounded GLR parser stack splits and heavy allocation churn.
• v0.24.0 / v0.26.0: Fixed parser hangs by introducing bounded GLR stack depth limits and ABI v14 state safety caps.
• v0.40.0: Added memory worker bounds and initial scratch buffer allocations.
• v0.48.0: Added O(N) leaf node matching, buffer reuse, and 200+ pre-compiled pure Go grammars for native TUI syntax highlighting.

Benchmark & CI Reporting Impact

• Python Parsing: Up to 17x faster on large Python files (py_fix_encode_files dropped from 392.57 ms to 23.31 ms).
• Heap Allocations: Cut Go heap allocations significantly (-35% to -53% fewer allocations across Go fixtures).
• Executive Summaries: PR CI runs render structured Markdown tables for benchmark comparisons and code coverage right in the GitHub summary tab when performance-critical paths are modified.
• CGo Evaluation: Confirms keeping pure Go gotreesitter with CGO_ENABLED=0, giving us zero CGo build setups with only ~10% total pipeline difference against CGo tree-sitter.