perf(analyzer): index qualified reference lookups

[mcdigman]

Closes #768

Summary

• index non-import definitions by simple name and source file
• resolve qualified references with logarithmic binary searches instead of repeated candidate scans

Knocks down the empirical O(n^2) scaling.

Measured Performance Improvement

• homeassistant/core/homeassistant: _mark_refs cumulative time: 1213.7 s -> 84.5s
  • direct str.startswith calls from _mark_refs: 2.661 billion -> 1.965 million
• scipy: 23.1s -> 14.3s in _mark_refs (scipy specifically is not as badly hit by this particular quadratic scaling relative to its size as a repo)
• Not much change in timing (maybe ~5% improvement within ~inter-run noise) on skylos's current SHA pinned liveness_primer corpus (small to medium sized repositories not as affected by quadratic scaling)

Note: I doubt this completely maxes out possible perf wins in _mark_refs; I thought the maintainer might find a smaller footprint/relatively straightforward code change from me that specifically targets the O(n^2) scaling with indexing more palatable than a larger refactor, given how cognitively complex and important this function is.

Validation

• Home Assistant findings were byte-for-byte identical across both 2,218-line outputs
• 188 analyzer tests passed
• 85 dead-code and framework-aware tests passed
• targeted Ruff undefined-name, unbound-name, and closure-binding checks passed
• liveness_primer shows 0 diffs on corpus

AI Use Disclosure
Profiled targeted and and potential O(n^2) identified by hand using cProfile. Codex implemented an indexing scheme and the sorting/bisection. Reviewed according to skylos repository review guidelines using GPT-5.6 Sol and Opus 5.0. Second commit was partly based on my own by-hand simplification findings.

[duriantaco]

LGTM. Thanks @mcdigman