Cache RefResolver subschemas lookup per instance

[sir-sigurd]

Issue #, if available: none — follow-up to #4586, found while profiling for #4585

Description of changes:

RefResolver._get_subschemas_cache() re-ran a full breadth-first walk of the entire referrer schema document (_search_schema) on every resolve() call, despite the name suggesting a cache. The referrer schema is immutable for the lifetime of a resolver instance, so memoize the result on the instance (a new init=False field).

Benchmarks (v1.53.2 baseline): on a ~330 KB real-world template, cumulative profile time in _get_subschemas_cache drops 1.65s → 0.05s (the top self-time hotspot remaining after #4586); end-to-end wall clock on that template goes 4.20s → 3.90s (Apple M-series, best of 6, ~−7%). On a CI runner, a same-runner interleaved 3-way comparison over 72 real-world templates measured ~−4.5% end-to-end with zero overlap between arms — full table in the comment below.

Full unit test suite passes (2644 passed, 1 skipped); ruff, ruff-format, isort, and mypy are clean on the changed file.

---

Prepared with AI assistance (Claude Code).

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[sir-sigurd]

Additional measurement in a CI environment, since Apple-Silicon numbers proved unreliable for small effects elsewhere: same-runner interleaved 3-way A/B/C (ubuntu-latest; 72 real-world templates, ~250–340 KB each; one full-set cfn-lint invocation per pass; 3 passes per arm, Latin-square ordering; venv contents assertion-verified).

arm	passes	min
1.53.2 (PyPI wheel)	316.4 / 317.9 / 316.2 s	316.2 s
1.53.2 base commit (git install — control)	320.0 / 325.1 / 315.9 s	315.9 s
this branch (git install)	301.5 / 310.8 / 305.3 s	301.5 s

Every pass of this branch is faster than every baseline pass: ~4.5% end-to-end on this workload — somewhat below the ~7% single-template number in the description, but cleanly separated from noise (the wheel arm's own spread was 0.3%).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.42%. Comparing base (928bbff) to head (e437cbb).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #4590   +/-   ##
=======================================
  Coverage   94.42%   94.42%           
=======================================
  Files         429      429           
  Lines       15099    15103    +4     
  Branches     2916     2917    +1     
=======================================
+ Hits        14257    14261    +4     
  Misses        461      461           
  Partials      381      381           

Flag	Coverage Δ
unittests	94.42% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.