add as_inexact option to _lib._util._asarray_validated

[argriffing]

In general, SciPy linear algebra functions prefer input arrays that don't contain NaNs, that are not object arrays or masked arrays, and that are dense (not sparse). Currently _asarray_validated checks all of those properties. This PR adds the additional option to convert the input array to a dtype that is more palatable to BLAS/LAPACK functions, and it should address one of the complaints in #4866.

[ev-br]

Maybe add a bit more flexibility by adding a dtype kwarg instead: _asarray_validated(..., dtype=None). Then dtype=None would mean to keep the input dtype (just like as_inexact=False is doing here), otherwise use toarray(.., dtype=dtype)?

[argriffing]

@ev-br There is some discussion about that kind of thing at numpy/numpy#5634, especially regarding extensions for flexibility.

dtype=None would mean to keep the input dtype (just like as_inexact=False is doing here), otherwise use toarray(.., dtype=dtype)

As I understand it, this suggestion would need at least some modification. For example the code in this PR is intended to convert int but not complex to float when as_inexact=True.

The TravisCI tests are failing because of test timeouts and also because the code is incompatible with supported numpy 1.6.
asarray_chkfinite() got an unexpected keyword argument 'dtype'

[ev-br]

Ah, OK. Indeed, it might be better to keep it simple for now and delegate additional sophistication to numpy. The numpy 1.6 limitation is I guess easy to work around with a try... except TypeError: (..).astype(np.float_) --- unless we want to bump the minimum numpy version again for scipy 0.17 :-).

[ev-br]

OK, this should not fall through the cracks. I've tested it locally and it seems to do what it says on the box:

>>> a
array([1, 1, 1])
>>> a.dtype
dtype('int64')
>>> from scipy._lib._util import _asarray_validated as av
>>> av(a, as_inexact=True).dtype
dtype('float64')
>>> 
>>> b = np.ones(3, dtype=np.complex128) * 2j
>>> b.dtype
dtype('complex128')
>>> av(b, as_inexact=True).dtype
dtype('complex128')
>>>
>>> c = np.ones(3)
>>> cc = av(c, as_inexact=True)
>>> np.may_share_memory(c, cc)
True

Until numpy grows the relevant functionality, this should do the it (and we'll see how much usage this actually gets).
Merging, thanks @argriffing.