checksum: add FxHash and GxHash based checksum

[fia0]

This is just to test out the performance compared to the xxhash we've been using until now. Early measurements with 4M blobs have shown that it could be worth experimenting with fxhash (https://lib.rs/crates/rustc-hash) which is used in the rust compiler. Another one worth trying out is ahash (https://lib.rs/crates/ahash) used by hashbrown. Ahash is too unstable.

[fia0]

Due to stability concerns i've tested some of the hash impls with a bunch of different compiler versions on different CPUs and indeed ahash is quite unstable, but fxhash remained remarkably stable. Included in the tests are ants epyc7000 (CentOS Stream 8), epyc8000 (CentOS Stream 8), epyc9000 (Rocky 9), xeon gold 5220R (Ubuntu 20.04), a virtual CPU with some xeon skylake below (Fedora 39), a virtual CPU with a neoverse-n1 below (rocky 9) and the work station of yours truly i7 13000 (Fedora 39). They all produce the same hash for a given input under multiple stable and nightly.

[fia0]

For a comprehensive overview refer to https://github.com/rurban/smhasher.

[fia0]

Numbers from one of the EPYC 9000 servers (hashing a randomly filled 4MiB slice):

GxHash:
	Minimum: 45441 ns
	50%: 46400 ns
	95%: 47980 ns
	Maximum: 53160 ns
ahash:
	Minimum: 123610 ns
	50%: 126410 ns
	95%: 128561 ns
	Maximum: 135941 ns
XxHash:
	Minimum: 217851 ns
	50%: 221981 ns
	95%: 226151 ns
	Maximum: 246701 ns
Highway:
	Minimum: 320102 ns
	50%: 324492 ns
	95%: 327932 ns
	Maximum: 344701 ns
FxHash:
	Minimum: 708773 ns
	50%: 710374 ns
	95%: 711574 ns
	Maximum: 727214 ns
ZwoHash:
	Minimum: 708794 ns
	50%: 714243 ns
	95%: 717874 ns
	Maximum: 728874 ns
Murmur3:
	Minimum: 3144406 ns
	50%: 3149066 ns
	95%: 3153416 ns
	Maximum: 3623948 ns

[fia0]

The PR should be mostly finished now, I'll perform some tests and then merge it down. From the hashing implementations now in the stack the order in regards to performance seems to be: GxHash > XxHash ~ FxHash. The latter two really are system-dependent, I've had systems in the testbed in which XxHash outperformed FxHash by 2x and some in which FxHash is somewhat faster than XxHash. Although overall, GxHash seems to be No. 1 in each of these machines.

[fia0]

The picture is similar for Highway. Which has a mixed performance overall depending on the system.

[fia0]

After some benchmarking I can validate the performance difference also in Haura, the differences are most notable when cache pressure is highest, for example on point queries. The figure below shows the performance scaling with XxHash in blue and GxHash in orange.

[fia0]

Based on the benchmark results, and since there is no data stored beyond ephemeral development stuff in all existing Haura deployments, I took the liberty to assign gxhash as the default checksum. The version for gxhash is pinned to v3.1.1.