Optimize the fit gradients

[HanatoK]

• Save the unrotated positions. This can avoid rotating the positions again in calc_fit_forces_impl;
• Determine dl, dq or ds by compile-time template options;
• Only call calc_fit_gradients when atom groups have explicit gradients;
• Simplify and inline cvm::quaternion::position_derivative_inner.

This PR should be only marked as ready after #713 is merged.

[giacomofiorin]

@HanatoK We don't currently have a CI job with Intel, but that compiler uses restrict (as well as PGI pre-NVIDIA). See the following from the LAMMPS headers:
https://github.com/lammps/lammps/blob/59bbc5bcc1104bdb4fb45107cd65b5d4d76dbc00/src/lmptype.h#L342-L348

[HanatoK]

@HanatoK We don't currently have a CI job with Intel, but that compiler uses restrict (as well as PGI pre-NVIDIA). See the following from the LAMMPS headers: https://github.com/lammps/lammps/blob/59bbc5bcc1104bdb4fb45107cd65b5d4d76dbc00/src/lmptype.h#L342-L348

Thanks! I will try using the same way to apply __restrict.

[giacomofiorin]

@HanatoK We don't currently have a CI job with Intel, but that compiler uses restrict (as well as PGI pre-NVIDIA). See the following from the LAMMPS headers: https://github.com/lammps/lammps/blob/59bbc5bcc1104bdb4fb45107cd65b5d4d76dbc00/src/lmptype.h#L342-L348

Thanks! I will try using the same way to apply __restrict.

Awesome! I added an Intel oneAPI CI job in the meantime, which is overkill for just this PR but definitely was a missing feature otherwise.

[HanatoK]

I think it is better to document how the fit gradients are computed so I tried my best to improve the doxygen documentation in 2d353eb. I know the SI of the new Colvars paper includes the computation but that is too simplified and not quite informative.

[jhenin]

Nice optimization I rebased and made tiny cosmetic changes, and I have one question. Apart from that, this is ready to merge IMO.

[giacomofiorin]

Looks great, apart from the speedup the code is also getting more readable.

[giacomofiorin]

Awesome, thanks!