turns out not only sparc is borked, skip the checks on all 32 bit arches
with too large clongdouble alignments until we have an aligned
allocator.

TST: Silence some warning that turns up on OpenBSD.

On OpenBSD with gcc-4.2 log1p(nan) raises an invalid value warning.
This PR disables both 'divide' and 'invalid' for the log1p tests.

Closes #5017.

TST: silence deprecation warnings for bad index calls

Some bad indexing calls throw exceptions and deprecations, silence the
warnings in the test.

BUG: set needs-init flag for unicode dtype

Python >= 3.3 does not like printing garbage unicode so printing
np.empty(x, 'U') can throw UnicodeDecode errors.
It is probably easier to just require initializing the dtype as a whole
instead of trying to catching the decode error on all places where it
might occur. Hopefully 0 is a valid code in all codecs.
Closes gh-4994.

ENH: np.dot: better "matrices not aligned" message

I had to move convert_shape_to_string to common.h so that _dotblas.c
can access it.