This was we don't get noisy baselines just because a different phase of the compiler reported
the diagnostic.
This helps with Yui's refactoring work to move grammar checks into the type checker.
We now no longer fail a speculative parse just because an error was encountered at any point while
speculating. instead, we allow the speculative function that is being called to determine if the
parse was successful or not. Only if it decides it was not successful is parsing rewound.
This improves our error recovery in several cases (esp. around arrow functions). it will also
help in a followup refactoring to prevent lookahead/speculative parsing from causing lambda
allocations.
