Faster GAF sorting

[jltsiren]

Changelog Entry

To be copied to the draft changelog by merger:

• GAF sorting with vg gamsort is much faster than before.

Description

New GAF sorting algorithm that does not parse the alignments beyond determining the key. It handles the records as pairs consisting of an integer key and a string value.

The initial sort splits the input into chunks of (e.g. 1 million) reads that are sorted into temporary files using worker threads. If the input is bgzip-compressed, 4 or 5 worker threads should be enough.

Intermediate merges take ranges of (e.g. 32) temporary files and merge them into larger temporary files using worker threads. This part parallelizes better than the initial sort.

Once the number of temporary files is low enough, the final merge is performed sequentially. If the output is bgzip-compressed, it may need 5 or 6 compression threads to avoid becoming the bottleneck.

The temporary files are compressed using zstd. This uses zstd_ifstream and zstd_ofstream, which are compressed versions of std::ifstream and std::ofstream with somewhat limited functionality.

With 5 worker threads and 5 bgzip threads, this should sort a 30x GAF in 15-20 minutes on a recent laptop.

There are also options for using a stable sorting algorithm and shuffling the reads instead of sorting them.

This PR also updates GBWT and GBWTGraph for more efficient GBZ loading.