Speed up and simplify SR

Created by: inailuig

This implements the non-centered Sv = ⟨Ô†ΔÔ⟩v with just 1 jvp + 1 vjp, instead of 1 jvp + 2 vjp all while doing the same amount of communication. Results in a measurable speedup compared to before and also compared to the centered version (benchmarking just the matrix-vector product).

Also fixes and simplifies a few things and extends the test.

A derivation is coming soon™.