Tools - mkbuildoptglobals refactoring & attempt to fix caching

[mcspr]

Optional include for build-dir header to exclude it from core.a when no build opts are created / used. Preventive reads before writes, too
Prefer to use --output-dir in scripts, tests/device must be out-of-tree for caching to work properly (by avoiding changing core nested dirs)

Assume 'aggressive caching' is normal mode of operation, no need for special treatment
(dependencies can be checked by looking at .d files in the build dir)

Even CI uses arduino-cli ~/.cache/arduino/{sketch,core} for building
Allow sketches without globals.h to share core.a, rebuild otherwise (same sketch cache is retained, core.a & .o's reused)

Drop manual logging in favour of 'import logging'. Arduino-CLI might still need tweaks in debug mode, though. Output flushing does not always happen promptly, but using single lines instead of multiple often helps.

Allow multiple entries for the same name, which are then merged.
Allow spaces between @ and the signature words.
Allow to use line right after signature.
Syntax warnings show relevant snippet from globals.h

[mcspr]

Also, ping @mhightower83

Main difference is .d dependency. Previously it was both CommonHFile.h & build-path/SKETCH globals.h. Now, it is either one or both. Sketches without globals.h should build as usual, no dependency means core.a is cached and no core files are depending on a different sketch build path every other build

globals.h is also included only when build.opt includes it. Direct dependency on the sketch can be created via #include "SKETCH.ino.globals.h", if it is ever necessary
(for our files, it is not. maybe someone actually uses it 🤷)

I kept previous 'aggressive caching' notes for posterity, since I am not sure how arduino-builder / arduino-cli worked (or did not work) before. Could also just remove it, since none of the notes actually apply to the current version of the script

w/ PR - IDE 1.8.19 works fine, so is CLI 1.2.2 (CI runs it)