Tests do not work on i386 due to use of float128

[jeremysanders]

On the Linux i386 platform (at least), the data type numpy.float128 does not exist. This causes the test test_cost.py to fail. I believe float128 is being used to map to long double in C++. numpy/numpy#10288, suggests that the data type numpy.longdouble should likely be used instead of float128, as long double is not an ieee128 datatype on that platform.

See https://gist.github.com/jeremysanders/9aca53cfccd5414d43f9f33eae911287 for a patch.

[HDembinski]

Thank you for this bug report, but I think the correct treatment is to skip this test on platforms that do not support this data type. Is there a way to check whether numpy supports numpy.float128 that can be used as a requirement for the test?

[jeremysanders]

I don't agree, as the test works anywhere fine if the data type is changed. The C data type is long double, so why not use the correct data type in iminuit? There's no guarantee that a long double is a float182. In addition, the presence of float128 doesn't mean that a long double has this same data type.

[HDembinski]

The test is explicitly for the type float128. A long double is not guaranteed to have 128 bits.