Build a weight=None from_networkx from the adjacency alone - #645

A structural conversion (weight=None) of a simple graph stores 1 for
every edge, yet from_networkx built it like a weighted one: a (u, v, 1)
tuple per edge through G.edges, three tuples transposed out of those,
an array of ones inferred from a Python sequence, and a uniformity test
over it before the iso build.

Read only the coordinates instead: for each node in nodelist order, its
neighbors straight from the adjacency, which for an undirected graph
already holds both directions and each self-loop once, so nothing is
mirrored. Rows come from np.repeat over the neighbor counts and columns
from one np.fromiter; Matrix.from_coo with the scalar 1 stores the
result iso on SuiteSparse and broadcasts it elsewhere. The dtype is what
the general path infers (INT64, or dtype= when given). An edgeless graph
still takes the general path, which gives it scipy's float64; so do
multigraphs, whose parallel edges sum, and a dtype= that is not bool,
int, float or complex.

gnm_random_graph(100000, 500000) plus one self-loop, best of 5 in one
warm process, 18 SuiteSparse threads, load average about 1.2:

    undirected  159.5 ms -> 75.0 ms  (2.1x; 199.4 ms before the self-loop fix)
    directed    131.8 ms -> 49.3 ms  (2.7x)

A nodelist covering half the nodes goes through a subgraph view and
gains nothing (178 against 180 ms undirected), but loses nothing either.

The new test compares weight=None against an attribute no edge carries,
which takes the general path and fills every entry with the default 1,
over non-integer keys, a self-loop, reciprocal edges, an isolated node,
nodelist reorders and subsets, an edgeless graph and three dtypes.
Over 124 conversions (every case of a parity script also with
weight=None) the result, dtype and iso flag match the parent's exactly.