Cache StackConfig loads keyed on file identity

[nicoloesch]

Summary

• Fixes Cache StackConfig loads to cut redundant disk reads #16

load_stack_config() re-read and re-parsed the config TOML from disk on every call, uncached, including from PackageConfigBase.get_config() and .get_engine(), which every consuming package goes through. Adds a process-local cache, keyed on (resolved_path, mtime_ns, size, $OA_ACTIVE_PROFILE, $OA_CONFIG_PATH), so any
file edit or relevant env var change invalidates automatically. No caller needs to invalidate anything for a normal read. Reuses the stat() call already made for the permission-mode check, so this costs zero extra I/O.

Every read, cached or not, still returns a fresh model_copy(deep=True), so the public contract is unchanged: callers get an independent, freely-mutable StackConfig exactly as before (verified against groundworkers, which
mutates the returned object directly). save_stack_config() and patch_active_profile() (the only two functions that write the file) clear the cache immediately after writing.

Checklist

• Applied exactly one label (feature)
• Tests pass locally (uv run pytest -q)
• Lint passes (uv run ruff check .)