Increase the default nonbonded cutoff from 7.5 A to 9 A

[lohedges]

Profiling a 46646 atom protein-ligand system with nsys showed that at the 7.5 Å default the PME reciprocal space kernels (charge spreading, FFTs and force interpolation) take about 2.5 times as long as the direct space kernel, because the short cutoff forces a large Ewald alpha and a fine grid. With PME the electrostatics are independent of the cutoff for a given error tolerance, so the cutoff is a numerical choice that moves work between the grid and the pairwise kernel; only the Lennard-Jones truncation changes, and since we apply no dispersion correction a longer cutoff is if anything more accurate. Timing an MD step with the production settings (HMR, 4 fs, langevin_middle, h_bonds constraints, barostat every 25 steps) on an RTX 5070 Ti gave 0.379 ms at 7.5 Å, 0.359 ms at 8 Å, 0.352 ms at 9 Å and 0.352 ms at 10 Å, with an energy evaluation going from 0.329 ms to 0.297 ms at 9 Å, which matters for the energy scan in replica exchange. The exact optimum depends on the GPU, but 7.5 Å is short for PME on any recent card, and 9 Å is what our users already run, so the default now matches. Results will not be bit-comparable with runs at the old default, as the Lennard-Jones cutoff and the Coulomb shift change with it.