* Rework IR and resolver to use nested lookup tables
* Integrate types into IR
* Make components prefer IR, slimmed down AST
* Implement `export *`
* Add `@lazy` annotation and remove `--noTreeShaking`
* Add `@start` annotation and remove magic `main`
* Related refactoring, cleanup and docs
This basically flips the --noDebug flag to become a --debug flag, so some optimizations, like inlining, aren't skipped by default, which might be unexpected.
Looking at our existing tests it seems that inlining these can be performed by reusing locals most of the time, and sometimes enables erasing quite a bit of unnecessary code when optimizing.
When specifying -O3 or -Oz, asc now automatically increases the optimize level to 4, then including costly non-LLVM optimizations for blocky code, in turn achieving similar results as an LLVM-based generator.
Remaining implementations of JavaScript's Math functions (except sin/cos/tan), both double (Math) and single (Mathf) precision, ported from musl incl. tests from libc-test, plus some changes to match JS semantics. Also binds fmod to `%` and pow to `**`.
