BLD: add -std=c99 flag to scipy.special extensions using cosine.c

[rgommers]

This fixes the manylinux1 builds. Closes gh-13840

Also clean up an outdated try-catch.

[rgommers]

Still a 32-bit Windows failure. This code is quite fishy:

    // M_PI is not exactly π.  In fact, float64(π - M_PI) is
    // 1.2246467991473532e-16.  h is supposed to be x + π, so to compute
    // h accurately, we write the calculation as:
    h = (x + M_PI) + 1.2246467991473532e-16;

Not only should you not rely on M_PI being available, but if it is available, the exact numerical value is going to be a compiler-dependent approximation to \pi. It's very hard to believe that all implementations are going to be off by about np.finfo(np.float64).eps - 1e-16.

I'd prefer to mark this as a knownfail and open an issue to follow up (if desired, it's quite a niche failure). @WarrenWeckesser can you have a look at this?

[rgommers]

That hexadecimal representation fixed it, merging. Thanks @WarrenWeckesser

[ev-br]

Maybe upstream it to npy_math.h's definition of NPY_PI ?

[rgommers]

That could make sense. Would you want to do that @WarrenWeckesser?