perf: memoize import resolution

[longlho]

Summary

• memoize per-analysis import resolution results inside ModuleResolver
• refresh benchmark docs for VS Code, Grafana, Kibana, and Renovate
• keep file reads and parsing unchanged; the optimization targets repeated resolver work rather than the FileCache path

Benchmark Notes

--samples 3 --warmups 1 --skip-build with the optimized Codescythe binary:

Fixture	Codescythe	Knip
VS Code	1111.9ms +/-1.36%	4223.1ms +/-4.89%
Grafana	833.2ms +/-4.11%	9513.4ms +/-56.75%
Kibana	12963.8ms +/-25.85%	53327.5ms +/-33.69%
Renovate	154.5ms +/-4.79%	900.5ms +/-6.09%

Kibana-specific direct warm timings improved from roughly 13.1s-13.2s before the resolver cache to 11.2s-12.5s after the cache in repeated local runs.

Validation

• git diff --check
• bazel build -c opt //crates/codescythe_cli:codescythe
• bazel test //crates/codescythe:codescythe_test
• bazel test //benchmarks:kibana_fixture_test

//benchmarks:kibana_conformance_test still fails locally only on the known existing macOS/Linux snapshot total-file count drift: local macOS generates 85952/85936, while the checked-in Linux CI snapshot is 85960/85944.