Convert from networkx without the scipy detour - #600

A networkx multigraph fell back to scipy in from_networkx because the
direct path could not sum the weights of parallel edges. The directed
branch now passes dup_op=plus when the graph is a multigraph, so
parallel edges accumulate exactly as scipy's coo -> csr summation does.
Simple graphs keep dup_op=None and are unchanged, and the undirected
branch already summed because it uses plus for the self-loop diagonal
correction. Numeric multigraphs no longer need scipy at all; non-numeric
weights still defer to it.

Parity with the retained scipy path holds for MultiGraph and
MultiDiGraph with parallel edges, parallel self-loops, reciprocal
parallel edges, parallel edges whose weights cancel to zero (both sides
keep the explicit zero), absent weight attributes (nx defaults to 1),
weight=None, bool and int weights, and nodelist permutations and
subsets. New tests pin the multigraph diagonal sum and assert the
numeric path never reaches the scipy fallback.

…ions

Converting through scipy ended in ``from_scipy_sparse``, which checks whether
every value is the same and imports one value instead of nvals of them. The
direct path added here ends in ``from_coo``, which does not, so an unweighted
graph quietly started storing a full values array.

Nothing caught it because SuiteSparse was still doing its own post-build iso
check, and the result came out iso either way. 10.5.0 dropped that check, at
which point ``test_matrix_to_from_networkx`` failed and the regression became
visible. The memory it costs was there the whole time.

Ask for iso outright, under the same uniformity test ``from_scipy_sparse``
uses. Multigraphs are excluded because ``plus`` sums parallel edges and can
make uniform inputs non-uniform. Undirected graphs need no exclusion: without
self-loops their coordinates are unique so ``plus`` never fires, and with them
the diagonal correction puts both ``wt`` and ``-wt`` in the values array,
which fails the uniformity test on its own.

The widening added while stabilizing CI was marked "pragma: no cover
(win64 numpy < 2)", but the branch is reachable on every platform:
np.int32 scalar edge weights infer an int32 array from the edge sequence
anywhere, not only where python ints do. The pragma therefore hid live
code from coverage, and the comment implied the change was platform-local
when it is user-visible everywhere: int32-weighted graphs that previously
round-tripped through scipy as INT32 now come back INT64.

The widening itself is kept, deliberately. from_coo widens an int32 dtype
inferred from sequence input to int64 on every platform
(values_to_numpy_buffer), and from_networkx builds its values from a
Python sequence, so identical weights now produce the same dtype through
either constructor. Gating the widening to win64 instead would have
re-introduced a from_coo/from_networkx disagreement everywhere else.

Drop the pragma, restate the comment in terms of the real contract, and
pin the behavior: np.int32 weights widen to INT64 and agree with from_coo,
an explicit dtype= is preserved exactly, and np.int16 stays INT16 through
both constructors. Release-note item: the from_networkx result dtype for
int32-weighted graphs changes from INT32 to INT64.

The direct ingestion path symmetrized an undirected graph the way
nx.to_scipy_sparse_array does: it mirrored every entry, self-loops
included, and appended -wt for each self-loop so that plus would sum the
diagonal triple wt + wt - wt back to wt. That arithmetic had four
consequences:

- The iso test added in a1fa8f3 reads the values array, which then held
  both wt and -wt, so an unweighted (or uniformly weighted) undirected
  graph with a self-loop was stored with a full values array. Through
  scipy it was stored iso, because from_scipy_sparse tested the summed
  values. Code that branches on A.ss.is_iso (graphblas-algorithms runs
  BFS instead of Bellman-Ford for an iso graph) took the slow path.
- An infinite self-loop came back nan (inf + inf - inf), and a finite
  one near the float max came back inf.
- An unsigned dtype= raised OverflowError under numpy 2, which cannot
  write -wt in it.
- bool weights came back INT64 when the graph had a self-loop (-True is
  the int -1) and BOOL when it did not.

A self-loop is its own mirror, so mirror only the off-diagonal entries.
Every stored value is then an edge's own weight, and a simple graph has
no repeated coordinate: it builds without an accumulator, and a uniform
one takes the iso path whether or not it has self-loops. Multigraphs
still sum parallel edges and parallel self-loops with plus. The second
pass over the graph for nx.selfloop_edges goes away.

nx.to_scipy_sparse_array still sums wt + wt - wt, so from_networkx now
differs from it in exactly those four cases. Compared with the parent
over 74 graphs (empty, edgeless, isolated nodes, nodelist subsets and
reorders, non-integer keys, int, float, bool, complex, int32, missing
and nan weights, multigraphs, and the error cases), nothing else changes.

The old diagonal correction negated each self-loop's weight. For an
np.uint8 weight that wraps (-np.uint8(1) is 255), and beside a Python
int weight the values infer int64, so the self-loop of weight 1 summed
to 1 + 1 + 255 = 257, with only a RuntimeWarning. The parent commit
fixed this along with the four cases its message lists; pin it too.

from_networkx checks a nodelist as nx.to_scipy_sparse_array does, so
that the direct path raises its errors without needing scipy, but no
test gave it an empty nodelist, a node not in the graph, or a duplicate.
Check each message, and that a missing node is reported before a
duplicate.