Hamiltonian replica exchange

[mcwitt]

Adds a generic implementation of Hamiltonian replica exchange (HREX) and a specialized version for use with RBFE implemented using the former.

Notes

• This implements a variant of nearest-neighbor HREX (described in 1) that performs many consecutive nearest-neighbor swap attempts per iteration to generate effectively uncorrelated permutations. The matrix of energies by replica and state is computed only once per iteration, so the cost is amortized over many neighbor swap attempts.
• We use the heuristic $K^3$ for number of neighbor swap attempts per iteration, where $K$ is the number of states. This is derived from the empirical bounds of $K^3$ to $K^5$ in 1.
• This does not yet switch to using HREX as the default method for production RBFE calculations (i.e. run_leg functions).

To do

• Add benchmark comparing RBFE performance with and without HREX enabled
• Avoid modifying global PRNG state
• Verify / tune added nightly tests

Example plots

Test system: evenly-spaced overlapping Gaussians

[maxentile]

🚀 -- nice!

Only nits in this pass, will attempt some tire-kicking before completing review.

[mcwitt]

Test expansion request: can we add an HREX workload to the performance benchmarking script test_benchmark.py?

Not a full free energy calculation, but running a solvent leg and complex leg with a couple settings of n_windows (maybe 10, 20, 50) for about 30 seconds - 1 minute each? Would be helpful to keep an eye on relative costs of swap attempts vs. MD.

@maxentile I added a nightly benchmark in 33f0e72.

This currently just prints ns per day estimates for sequential and HREX simulations for vacuum, solvent, and complex legs with 5, 10, and 20 windows. (More than 20 is difficult to run in a reasonable time with the restriction that we keep the ratio of MD to HREX swap moves constant)