Avoid spread intrinsic

[foxtran]

With gfortran, spread intrinsic is not optimized (more precisely this is libgfortran call) and therefore it significantly affects on performance, especially for 3-body terms.

This patch speeds up gradient code in 1.5-2x times with gfortran as well other parts which I did not measure. By some reason, it also affects ifx but it gives only 10% speed up for deriv_atm_triple subroutine.

I have used unrolled cycles since they provide an extra optimizations for ifx: time was improved from 680 to 625 ms for my input. Usual cycles does not give this improvement: see the assembler difference here: https://godbolt.org/z/zs6e7vzT4. The first source code is the original, the second is presented in this PR and the third one uses cycles.

I did not touched initializations/io parts where spread are used.

[thfroitzheim]

I did not touch spread hell in gfnff/gfnff.f90. Also, I did not touched initializations/io parts where spread also used.

I know this is the annoying part, but GFN-FF would likely benefit the most from improved parallelization (i.e. in #1240). Likely much more than just a few ms here and there

[foxtran]

Likely much more than just a few ms here and there

It is not a few ms for here and there for GCC, unfortunately.

[foxtran]

For GFN-FF and bench.xyz/1666_0064xGluAla.xyz, for GCC I got a 20% speed-up with this patch: from 0.78 sec/iter to 0.63 sec/iter in single-threaded run.

[foxtran]

@thfroitzheim, ping :-)

[thfroitzheim]

LGTM