Add fast_float header library for fast string->double, and optimize Parquet WKT->WKB conversion

[rouault]

Background for this work: opengeospatial/geoparquet#170

Several related items:

• add a vendored copy of https://github.com/fastfloat/fast_float in third_party/ subdirectory.
• add a port/include_fast_float.h that uses the system fast_float/fast_float.h if found (determined by __has_include("fast_float/fast_float.h") or by building explicitly with -DUSE_SYSTEM_FAST_FLOAT=1), or fallbacks to the vendored one
• update CPLStrtod() family of function to use fast_float
• Parquet: add a COORDINATE_PRECISION layer creation option to set the number of decimal figures to WKT geometries (when using -lco GEOMETRY_ENCODING=WKT)
• Parquet: when using the Arrow Array reader with the GEOMETRY_ENCODING=WKB option that forces geometries to be encoded in WKB, add an optimized WKT->WKB translator that avoids going through intermediate OGRFeature, and add a very specifc optimization for translation of single-part single-ring multipolygons, as used by the "reference" https://storage.googleapis.com/open-geodata/linz-examples/nz-building-outlines.parquet dataset which contains 3.2 million features with 13 attributes each and polygons that have generally a small number of vertices each (< 10).

Benchmarks:

• Creation of GeoParquet files (using Snappy compression):

Variant of geoparquet	Timing	File size
WKB	12.2 s	436 MB
WKT, 15 significant digits	1m 36s	666 MB
WKT, 4 decimals	1m 27s	474 MB

This shows that generation of WKT is very slow compared to WKB (I presume this could be optimized to be at least twice faster)

• Running bench_ogr_c_api, that is iterating over OGRFeature returned by OGRLayer::GetNextFeature()

Variant of geoparquet	Timing
WKB	5.9 s
WKT, 15 significant digits	14.1 s
WKT, 4 decimals	12.3 s

The timings of the WKT parsing have been improved through the use of the fast_float library (e.g. 16.8 s before for "WKT, 4 decimals" dataset when using strtod())

• Running bench_ogr_batch --stream-opt GEOMETRY_ENCODING=WKB, that is using the Arrow Array streaming mode, but forcing geometries to be WKB even if it is not their native encoding

Variant of geoparquet	Timing
WKB	0.96 s
WKT, 15 significant digits	5.3 s
WKT, 4 decimals	4.6 s

The timings with the WKT files have been considerably improved by avoiding the round-trip to OGRFeature and using fast_float as well (e.g 25 s before for "WKT, 4 decimals" dataset when going through OGRFeature and using strtod())

For those read benchmarks, we're probably clause to the optimal performance