The processing of long numbers could be a bit faster

[lemire]

Following https://github.com/lemire/fast_float/pull/15, we have honest performance in the more-than-19-digits scenario, but anyone reading the code paths will see that there is obvious (non highly technical) room from speed gains.

This is not a priority because beating speed records in that unusual scenario is not very important, and we are still typically going to beat standard libraries. We just want to avoid really bad performance.

The parse_decimal is subject to optimizations:

1. We do not need to reparse the exponent, we could recover it from our pass in parse_number_string.
2. We could stop after 19 digits and try to bail out using the fast slow path from https://github.com/lemire/fast_float/pull/15

If someone is looking for some fun work...

[lemire]

There is a prototypical PR at #27 but it adds about 5% of overhead to the general case, to the benefit of a specific case that is of little practical interest.

[lemire]

Another PR #51 with a different trade-off.