Fix performance regression when parsing bytes.

[rsmmr]

Turns out our improved error messages were adding additional overhead
because we were now constructing them through fmt() each time we
needed more data, independent of whether there was actually going to
be an error reported.

This adds a second version of waitForInput() that doesn't receive an
already prepared error message, but just returns false on error so
that the caller can throw itself.

Closes #1908.

[evantypanski]

Would it make sense to expand this more in the future so it doesn't come up again? There's a doc comment now, which is good, but more guarantees to prevent us from shooting ourselves in the foot could be nice in the future (not here though, probably)

Something like a lazy-evaluated fmt would be really cool. But probably not necessary :)

[rsmmr]

Would it make sense to expand this more in the future so it doesn't come up again?

Yeah, good point. In fact I was trying to come up with something that would enforce correct usage, but couldn't think of anything viable that wouldn't beed larger changes. My two ideas were: (1) leveraging the type system by passing a const char* so that the result of fmt() wouldn't be accepted, and (2) passing a lambda instead of the error message, similar to your idea of a lazy-evaluated fmt. But (1) broke other stuff with generated code passing string_views; and (2) I think would itself be too expensive because it would now need to capture each time. Let me know if you have other ideas!