ARROW-10328: [C++] Vendor fast_float number parsing library

[pitrou]

This library is 2x to 3x faster for parsing strings to binary floating-point numbers.

[pitrou]

Before:

• Parsing raw values:

FloatParsing<FloatType>          39763 ns        39758 ns        52350 items_per_second=25.152M/s
FloatParsing<DoubleType>         29812 ns        29808 ns        68765 items_per_second=33.5483M/s

• CSV converter performance:

FloatConversion                 199539 ns       199508 ns         3498 items_per_second=40.0986M/s

• Reading a CSV file of floating-point numbers (single-threaded):

276 MB (10000000 rows) in 1.481 s. => 187 MB/s.
276 MB (10000000 rows) in 1.441 s. => 192 MB/s.
276 MB (10000000 rows) in 1.417 s. => 195 MB/s.
276 MB (10000000 rows) in 1.418 s. => 195 MB/s.
276 MB (10000000 rows) in 1.430 s. => 193 MB/s.

[pitrou]

After:

• Parsing raw values:

FloatParsing<FloatType>          11936 ns        11934 ns       184367 items_per_second=83.7942M/s
FloatParsing<DoubleType>         10989 ns        10988 ns       194016 items_per_second=91.0101M/s

• CSV converter performance:

FloatConversion                 108180 ns       108166 ns         6442 items_per_second=73.9605M/s

• Reading a CSV file of floating-point numbers (single-threaded):

276 MB (10000000 rows) in 0.903 s. => 306 MB/s.
276 MB (10000000 rows) in 0.844 s. => 328 MB/s.
276 MB (10000000 rows) in 0.818 s. => 338 MB/s.
276 MB (10000000 rows) in 0.812 s. => 341 MB/s.
276 MB (10000000 rows) in 0.800 s. => 346 MB/s.

[github-actions]

https://issues.apache.org/jira/browse/ARROW-10328

[wesm]

Do we need to unvendor double-conversion?

[pitrou]

No, we still use it for the other way round (float to string).

[bkietz]

Nice speedup, I'll wait to merge until attribution has been resolved