CSR matrix should infer data type consistent with what COO does

[jni]

When using list inputs for (data, (i, j)), COO matrix infers the data type from the type of the items in data, just as np.array would.

In contrast, when constructing a CSR matrix with a (data, indices, indptr) tuple, either the data value must be an array, or the dtype= keyword argument must be passed to the constructor:

In [30]: i = [0, 1, 1, 1, 1, 2, 3, 4, 4]

In [31]: j = [2, 0, 1, 3, 4, 1, 0, 3, 4]

In [32]: data = [6, 1, 2, 4, 5, 1, 9, 6, 7]

In [33]: indptr = [0, 1, 5, 6, 7, 9]

In [34]: coo = sparse.coo_matrix((data, (i, j)))

In [35]: csr = sparse.csr_matrix((data, j, indptr))
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/Users/nuneziglesiasj/anaconda/envs/elegant/lib/python3.4/site-packages/scipy/sparse/sputils.py in getdtype(dtype, a, default)
    116         try:
--> 117             newdtype = a.dtype
    118         except AttributeError:

AttributeError: 'list' object has no attribute 'dtype'

During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)
<ipython-input-35-8de58fac1a34> in <module>()
----> 1 csr = sparse.csr_matrix((data, j, indptr))

/Users/nuneziglesiasj/anaconda/envs/elegant/lib/python3.4/site-packages/scipy/sparse/compressed.py in __init__(self, arg1, shape, dtype, copy)
     54                     self.indices = np.array(indices, copy=copy, dtype=idx_dtype)
     55                     self.indptr = np.array(indptr, copy=copy, dtype=idx_dtype)
---> 56                     self.data = np.array(data, copy=copy, dtype=getdtype(dtype, data))
     57                 else:
     58                     raise ValueError("unrecognized %s_matrix constructor usage" %

/Users/nuneziglesiasj/anaconda/envs/elegant/lib/python3.4/site-packages/scipy/sparse/sputils.py in getdtype(dtype, a, default)
    121                 canCast = False
    122             else:
--> 123                 raise TypeError("could not interpret data type")
    124     else:
    125         newdtype = np.dtype(dtype)

TypeError: could not interpret data type

In [36]: sparse.csr_matrix?

In [37]: csr = sparse.csr_matrix((data, j, indptr), dtype=int)

In [38]: data2 = np.array(data)

In [39]: csr = sparse.csr_matrix((data2, j, indptr))

In [40]: coo.todense()
Out[40]:
matrix([[0, 0, 6, 0, 0],
        [1, 2, 0, 4, 5],
        [0, 1, 0, 0, 0],
        [9, 0, 0, 0, 0],
        [0, 0, 0, 6, 7]])

In [41]: csr.todense()
Out[41]:
matrix([[0, 0, 6, 0, 0],
        [1, 2, 0, 4, 5],
        [0, 1, 0, 0, 0],
        [9, 0, 0, 0, 0],
        [0, 0, 0, 6, 7]])

I would argue that CSR should infer the data type in the same way that COO does, even if the input is a list (or other array-like).

[argriffing]

should be fixed in #5354

[jni]

Wow. Speedy service! Thanks @argriffing! =D