This commit adds fuzzing for the (expr) formatter, with the same invariants that we use for fmt tests:
* We start with text, which we parse
* We format the AST, which must succeed
* We parse back the AST and make sure it's identical igoring whitespace+comments
* We format the new AST and assert it's equal to the first formatted version ("idempotency")
Interestingly, while a lot of bugs this found were in the formatter, it also found some parsing bugs.
It then fixes a bunch of bugs that fell out:
* Some small oversights in RemoveSpaces
* Make sure `_a` doesn't parse as an inferred type (`_`) followed by an identifier (parsing bug!)
* Call `extract_spaces` on a parsed expr before matching on it, lest it be Expr::SpaceBefore - when parsing aliases
* A few cases where the formatter generated invalid/different code
* Numerous formatting bugs that caused the formatting to not be idempotent
The last point there is worth talking further about. There were several cases where the old code was trying to enforce strong
opinions about how to insert newlines in function types and defs. In both of those cases, it looked like the goals of
(1) idempotency, (2) giving the user some say in the output, and (3) these strong opinions - were often in conflict.
For these cases, I erred on the side of following the user's existing choices about where to put newlines.
We can go back and re-add this strong opinionation later - but this seemed the right approach for now.
On my M1 mac this shows as ~25% faster at parsing Num.roc than the old implementation, probably because nobody wrote any NEON code.
Even on my x86_64 linux box (Ryzen 2700x), this shows as 10% faster than the current SSE implementation (running with RUSTFLAGS="-C target-cpu=native").
* The header + expr fuzzers can both be run again (header fuzzer had regressed).
* I ran the expr fuzzer for ~60 seconds with no additional panics uncovered
* "tab_crash" hit supposedly unreachable code in blankspace.rs - and I went to the liberty of dramatically simplifying all that code, rather than just trying to fix the bug
* Other failures were straight-forward error cases that should have been handled (and passed up the chain) instead of panicking
