Proposal: use 8 rounds of ChaCha by default

[kernelmethod]

Right now we use ChaCha20 for CRNG by default (i.e., when you call ChaChaStream()). This is fairly expensive compared to the default PRNG used by Julia (which uses Xoshiro256++).

It seems that Golang nowadays uses ChaCha8 for math/rand/v2 (golang/go#61716). Based on the discussion there, it'd probably be worth investigating whether we'd be better off using ChaCha8 instead of ChaCha20 for CRNG.

[kernelmethod]

Some more info:

• It looks like rust-random/rand now uses 12 rounds by default. See the discussion here: rand_chacha: consider ChaCha12 (or possibly ChaCha8) over ChaCha20 rust-random/rand#932
• That issue also cites the "Too Much Crypto" paper from RWC'19, which is worth considering: https://eprint.iacr.org/2019/1492

[kernelmethod]

Final reference: https://go.dev/blog/chacha8rand

I've gone ahead with making ChaCha8 the default in the dev branch, which should provide an amortized ~2.5x speedup over the previous default of ChaCha20.