Should we add a series of bluk column decompress benchmarks

[joseph-isaacs]

e.g. string, int, float plus a few projections

This will test raw vortex decompression speed

[danking]

Generate 10,000 random normal 1400-d vectors, write them to a file, scan the file, call read_all().to_canonical().

This would be great! You could clamp them to some exponent range if you wanted to encourage ALP-RD.