This is a bit more then it used to do, so it is not a complete
revert. Some of the "weird" cases where a copy was unnecessarily
done will now only be gone with RELAXED_STRIDES_CHECKING.

Note that using Newshape also means that less copying is done
in principle, because ravel will always return a contiguous
array.

REL: Reset strides for RELAXED_STRIDE_CHECKING for 1.10 release.

Timedelta multiplication and division relied on undefined behaviour for some
inputs, which would give the expected results on x86 but not on armv7l
(e.g. Raspberry Pi 2).  Closes #6274.

Regression introduced in 4b1f508a57549d8031a23160b40c7f87f47892ed

Closes gh-6250.

BUG: simplify code remove strict aliasing violation

Didn't think the violation could cause issues but apparently some
compilers do mess it up. Also the old code is overly complicated, don't
know what I was thinking ...

Closes gh-6251

npyiter can change its inner stride when it is buffered, this
happens in some cases (the test is one of these) mostly for weirdly
strided arrays. The test does depend on the buffer size being used.

closes gh-6221

PyObject_HashNotImplemented was being used to initialize the tp_hash
slot in the PyArray_Type structure. In mingw32 (gcc version 3.4) that
results in a compile error. Fix by initializing to zero, then
setting to PyObject_HashNotImplemented when the module is loaded.

The function initialization seems to work with more recent compilers, as
the determination of what is considered 'non-constant' is left to the
vendor and has changed.