perf(runner): cache data-provider lookup per file

[Chemaclass]

🤔 Background

Related #763

Every test paid a grep -B2 | sed fork just to check whether it had a @data_provider annotation, re-reading the file from disk once per test. The scan is identical for every function in a file, so it only needs to happen once.

💡 Changes

• Scan each file's provider annotations once into a memoized fn -> provider map; per-test lookups are now pure-bash (no fork). Same single scan removes the per-test capture fork in both the run path and the test-counting path.
• Fork census on a 100-test file: grep 202 -> 2, sed 300 -> 100, awk ~1 per file (memoization confirmed, not per-test); ~14% faster wall-clock (load-variable). Win is platform-independent, unlike the clock work.
• Behaviour preserved: full suite 1218 -> 1225 (exactly the 7 new provider-map tests), green sequential and --parallel. A dropped total would have signalled missed detection. get_provider_data retained as a public wrapper over the new map.
• Repo-wide shfmt drift (v3.13.1 vs the version the tree was formatted with) was left untouched; the added code shows no drift and there is no shfmt CI gate.