Move choice of floating-point precision to the Python layer

[JonathanMaes]

This PR simplifies the process of switching between single and double floating-point precision.
Previously, the precision had to be chosen before compilation, by editing CMakeLists.txt.
Now, mumax+ compiles two binaries: one for single and one for double precision (unless MUMAXPLUS_FP_PRECISION is set, see below). The appropriate binary is then loaded during the import mumaxplus statement.

Two switches control the floating-point precision, as listed below in descending order of priority. Both can take the values SINGLE/1/32 for single precision and DOUBLE/2/64 for double precision.

• A command-line flag --mumaxplus-fp-precision passed to the python command.
  • Example: python some_mumax_script.py --mumaxplus-fp-precision DOUBLE.
• The environment variable MUMAXPLUS_FP_PRECISION.
  • Setting this will also affect the compilation: mumax+ will then only be compiled for one kind of floating point precision. This can be useful to reduce compile time during development, but passing the opposite precision with --mumaxplus-fp-precision will then result in a ModuleNotFoundError.

The used precision can then be accessed as mumaxplus.FP_PRECISION.

Note

Since the choice of precision must be made before the import mumaxplus statement, it is not possible to use single precision in one part of a script and double precision in another: these must be run in separate processes. Reloading mumax+ with a different floating point precision during runtime may lead to a vast wealth of errors.

The unit test is based on examples/errorscaling.py, since the error scaling is highly dependent on the floating point precision.

[ilateur]

Someone (or some-multiple) should test this on linux

[DiegoDeGusem]

On my linux device the code compiles and the test succeeds 🎉
I also agree with @ilateur's comment about a larger description.

[JonathanMaes]

A webpage was added in the Tutorials section, explaining how to choose a precision, what errors can occur, and showcasing the effect on the error floor.

[lamoreel]

I was able to successfully compile with a single (no pun intended) and with both precisions on a Linux machine 🌈 .
I was also able to choose the precision to be used by both the flag and the env var, hence my pending approval.

Most of my comments are about documentation

[lamoreel]

This doesn't work for me 😕 . (The kernel dies, but refuses to restart when trying to execute the next cell).

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
Cell In[1], [line 8]
        5 import numpy as np
        6 from math import acos, atan, pi, exp, tan, sin, cos, sqrt
  ----> 8 from mumaxplus import *
        9 from mumaxplus.util import *
       12 def magnetic_moment_precession(time, initial_magnetization, hfield_z, damping):

ModuleNotFoundError: No module named 'mumaxplus'

[JonathanMaes]

Looks like the kernel restarts, but not in the right conda env? The imports before import mumaxplus work just fine.