perf: optimize DWARF line and abbreviation lookups

[swananan]

Summary

• store compacted line rows and address groups in contiguous vectors while keeping an ordered incremental tail for lazy CU loading
• compact the incremental line index geometrically to avoid rebuilding all previously loaded rows on every CU query
• prepopulate gimli abbreviation tables once and reuse their shared Arcs across source-line, function, variable, and type lookups
• generate an in-memory DWARF fixture so cache coverage is independent of test-profile debuginfo

Performance

The abbreviation-cache step was benchmarked in isolation against the preceding line-index commit on the fixed query-hotspot corpus (12 alternating process pairs, 1,000 queries per process):

• average source-line query: 3.405 ms -> 1.643 ms (51.7% faster)
• first query: 5.335 ms -> 2.796 ms (47.6% faster)
• p95 query: 3.885 ms -> 1.951 ms (49.8% faster)
• parse-stress average load: 50.86 ms -> 50.17 ms (no observed regression)

Validation

• cargo fmt --all -- --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test -p ghostscope-dwarf (266 unit tests plus compatibility coverage)
• cargo test -p ghostscope-dwarf --release loaded_dwarf_reuses_cached_abbreviations --all-features
• full host-to-host e2e through the project runner (7024c13cd8f0)

Container-topology e2e was skipped because the changes do not affect containers, PID namespaces, or runner topology.