perf: probe toolchain once per daemon, not per TU #1466

Part of the cold-build perf meta issue.

strace of one cold Uno build shows 25× avr-g++ -v -E -x c++ /dev/null (system-include discovery), ~14× --version and 3× -dumpversion, when one of each per (compiler path, language, flags that affect includes) would do. Each spawns the driver plus cc1plus. Memoize them in the daemon, keyed by the compiler binary identity.

Acceptance: at most 1 probe of each kind per compiler per daemon lifetime, checked by a test that counts spawns.