BUG: Return NULL from PyInit_* when exception is raised

That function is nose specific and has not worked since `__init__` files
were added to the tests directories.

Numpy can now be tested using the standard

    `python -c"import numpy; numpy.test()"`

construct.

The "bench" testing with the old bench files is no longer supported.
These days we use `runtests.py` and `asv`.

I don't think this is documented anywhere, but I'm pretty sure module init
functions should return NULL in order to communicate that an exception
occurred during initialization (as is the standard Python/C API convention).

It's clear from the CPython code
[here](https://github.com/python/cpython/blob/master/Python/importdl.c#L162)
that if you don't return NULL, the exception is swallowed and replaced with the
message "initialization of %s raised unreported exception".

Admittedly, this is only useful for people porting Numpy to new platforms where
it is helpful to know where module initialization is failing, but it can't hurt.

MAINT: Stop using non-tuple indices internally

Previously, the numerical constants in the FFT code were not provided at full double precision which led to a loss of accuracy in the FFT operation. Additionally
this commit improves the accuracy of the twiddle factor calculation by reducing the argument of exp(2j*pi*m/n) to the first octant before calling the library
function. On average the commit lowers the remaining numerical error in the FFT by a factor of ten.

By not using this type of indexing, it becomes easier for subclasses to override indexing in a way that works correctly with numpy functions.

These locations were found by deprecating the behavior in question, which is deliberately not part of this commit

See the PR for benchmarking information