This should allow us to collect detailed information without slowing
down the inital hot path.

This moves out the matching part of expansion into a new function. This
function will try to match the macro and return an error if it failed to
match. A tracker can be used to get more information about the matching.

For now, we only collect the small info for the `best_failure`, but
using this tracker, we can easily extend it in the future to track
things with more performance overhead.

We cannot retry cases where the macro failed with a parser error that
was emitted already, as that would cause us to emit the same error to
the user twice.

These were useful while debugging, so I'll leave them here.