proper implementation of float and double parsers from text representation

[Geal]

right now we use recognize_float then call .parse::<f64>() on the recognized input, which means the data is parsed twice. It is necessary because parse() assumes it has the full input, and will not return remaining input data if there is some after the number. This is the main reason for serde being faster than nom on JSON benchmarks: serde constructs the number as it is parsing it.
We need a float parser that does that.

[pickfire]

What about parsing something like 0.01f32 as seen in rust?

[Geal]

@pickfire not really in scope. Anyone can implement that by writing terminated(float, tag("f32"))

[Geal]

nom 5 will now use the lexical-core crate to parse float since 40f20aa