Performance of deserializing IgnoredAny

[dtolnay]

I haven't benchmarked but there may be an opportunity to optimize IgnoredAny. Currently the codepath for deserialize_ignored_any is the same as for deserializing a serde_json::Value. It jumps back and forth into the IgnoredAny's Visitor and does all of the work that entails, including buffering up long strings, converting floating point values to f64, etc.

Instead we should be able to very quickly scan and do only the most basic syntactic checks, not buffering strings or interpreting numbers.

This could be a good intro task for somebody interested in performance. I can provide mentorship.

[dtolnay]

The Pikkr benchmark will be a good way to measure this https://github.com/pikkr/pikkr. They ignore most of the data.

[dtolnay]

I implemented this in #356. It looks like a nice 45% performance improvement on that benchmark.