ENH skip NPY_ALLOW_C_API for UFUNC_ERR_PRINT

[ndjensen]

GIL unnecessary when numpy floating point error handling is set to print
potential micro-optimization when error handling is set to print
alleviates (but does not fix) #5856 deadlock with sub-interpreters

Change-Id: I0414df8c5dca131e1f8c8d867791ba63cf992b63

[seberg]

Frankly, I am have not much clue about this, but the tests error out because you have to put the NPY_ALLOW_C_API_DEF before the statement (C89 compatibility -- -Werror=declaration-after-statement). While at it, can you could also remove the ; from it (it is prettier to put it, but it is already included in the macro, and ;; is bad because then you cannot add more defs after this defs for the same compatibility reason).
Anyway, generally seems right to me, not too pretty, but oh well....

[seberg]

I am sorry, I seem to have forgotten to mention that this also should use /* blah */ style comments, // are C++ style comments.
After that, please squash the commits, and I think we can merge it (might want to post a message, otherwise nobody notices the update), if you can think of a way to add a test, that would be great, but I am not sure there is anything not overly complex.

[seberg]

OK, looks good to me, seems weird to grab the GIL, if we have an explicit print option, so merging. Thanks.