Build musllinux wheels

[richli]

As of numpy 1.25, musllinux wheels are built and distributed on PyPI, supplementing the existing manylinux ones. It would be helpful for cftime to likewise build wheels for this platform.

It should essentially be a matter of updating the .github/workflows/cibuildwheel.yml to remove *musllinux* from the CIBW_SKIP variable. However, when I tried running it locally, the wheels built successfully but one test fails:

=================================== FAILURES ===================================
___________________________ test_num2date_precision ____________________________
                                                                                                                               
    def test_num2date_precision():           
        if np.finfo(np.longdouble).precision < 18:
            pytest.skip("skipping tests that require extended precision longdouble type")
        testdates = [(1271, 3, 18, 19, 41, 33),                                                                                
                     (1271, 3, 18, 19, 41, 32, 999998)]      
        unitinc = ['microseconds', 'seconds', 'minutes', 'hours', 'days']
        for cc in ['standard', 'gregorian', 'julian', 'proleptic_gregorian',
                   'noleap', 'all_leap', '365_day', '366_day', '360_day']:                                    
            for uinc in unitinc:             
                if cc in ['standard', 'gregorian', 'julian']:
                    units = uinc + ' since -4713-01-01 12:00:00'
                elif cc in ['proleptic_gregorian']:                                                                            
                    units = uinc + ' since -4714-01-01 12:00:00'
                elif cc in ['noleap', 'all_leap', '365_day', '366_day', '360_day']:
                    units = uinc + ' since 0000-01-01 12:00:00'        
                # scalar
                date = datetimex(*testdates[0], calendar=cc)                                                                   
                num = date2num(date, units, calendar=cc, longdouble=True)                            
                date2 = num2date(num, units, calendar=cc)                                                                      
                assert date == date2                                                                                           
                # array                                                                                                                                                                                                                                       
                date = [ datetimex(*dd, calendar=cc) for dd in testdates ]                           
                num = date2num(date, units, calendar=cc, longdouble=True)                  
                date2 = num2date(num, units, calendar=cc)
                for i in range(len(date)):                                                                                     
>                   assert date[i] == date2[i]                                                                                 
E                   AssertionError: assert cftime.datetime(1271, 3, 18, 19, 41, 32, 999998, calendar='standard', has_year_zero=False) == cftime.DatetimeGregorian(1271, 3, 18, 19, 41, 33, 0, has_year_zero=False)

/project/test/test_cftime.py:2212: AssertionError

It looks like numpy has had some similar test failures for long doubles using musl. This comment identified that on the x86_64 architecture, long double is actually the x87 80-bit float, not the IEEE-754 128-bit float type.

[jswhit]

I created PR #308 - and it shows the same test failure, only for Python 3.9 for some reason. It is true that long doubles are actually 80, not 128 bits on x86_64, but that doesn't explain why all the x86_64 tests pass except for musllinux with python 3.9.

[richli]

It also fails for Python 3.10 and 3.11, it's just that since 3.9 fails first it doesn't try testing those ones.

[richli]

I don't think it's a Python version issue, but a numpy version issue. With Python 3.7 and 3.8, it's using an older version of numpy.

[jswhit]

PR #308 merged, musllinux wheels should be created in the next release