This adds fused type to support 64 bit indices in csr_row_norms, inplace_csr_row_normalize_l1 and inplace_csr_row_normalize_l2. Fixes #6468 and begins to adress #2969 .

In this PR, for CSR array indices I define a fused integral type ( int, long long) which should produce either 32 bit and 64 bit respectively on Linux / Windows and 64bit / 32bit platforms as far as I understand. The array indices variables are always np.npy_intp, while the array shapes (M, N) are long long to support up to 64 bit array sizes. Please let me know if I missed something.

cc @jnothman @lesteve @ogrisel

Also tested that the performance is not affected (with 32bit CSR indices) using the following script (needs to be run with ipython),

import numpy as np
import scipy.sparse as sp
from IPython import get_ipython

from sklearn.utils.extmath import row_norms
from sklearn.utils.sparsefuncs_fast import (inplace_csr_row_normalize_l1,
                                            inplace_csr_row_normalize_l2)

ipython = get_ipython()
np.random.seed(43)
Xcsr = sp.random(10000, 10000, format='csr')

print('row_norm(Xcsr): ', end='')
ipython.magic("%timeit row_norms(Xcsr)")
print('inplace_csr_row_normalize_l1(Xcsr): ', end='')
ipython.magic("%timeit inplace_csr_row_normalize_l1(Xcsr)")
print('inplace_csr_row_normalize_l2(Xcsr): ', end='')
ipython.magic("%timeit inplace_csr_row_normalize_l2(Xcsr)")