Reimplemented alt error prioritisation, added more recovery strategies

[zesterer]

This is the last really big thing that needed implementing for zero-copy.

I was really worried that implementing it would hurt our benchmarks a lot since a very substantial proportion of the work that happens on master is related to error prioritisation.

However, I did some thinking and came to the conclusion that, for the most part, we can treat the returned error of a parser and the alt error as the same thing, massively simplifying its implementation and speeding things up a lot.

On the JSON benchmark, we only lose about ~7% with this change, still placing us waaay ahead of nom and friends. We pretty much regain the same high-quality error generation capabilities as are present on master, while also simplifying the implementation of parsers, for the most part.