Add PME parameter options and automatic PME tuning

[lohedges]

Profiling showed that OpenMM's default PME parameters often use a much finer reciprocal-space grid than the requested accuracy needs. Choosing them more carefully gives a decent speedup at the same or better accuracy. This PR uses the new Sire support to add pme_alpha, pme_grid and pme_spacing options for setting the PME parameters directly, and a pme_tolerance option for the Ewald error tolerance, whose default is unchanged. It also adds a tune_pme option, on by default. On the CUDA and OpenCL platforms it chooses the fastest PME parameters that are at least as accurate as those OpenMM would choose from pme_tolerance. Tuning runs once at setup, takes tens of seconds, and logs the chosen parameters and their relative force error. The result is saved to pme_parameters.yaml in the output directory and reused on restart, so a run keeps the same parameters even if it's restarted on different hardware. Restarts of runs from before this feature carry on with the default parameters. Tuning is skipped when explicit PME options are given, and on the CPU platform. This depends on OpenBioSim/sire#499.