Convert from networkx without the scipy detour - #600

from_networkx always routed nx -> scipy CSR -> Matrix, so every
conversion paid for scipy.sparse (importing it alone costs over 100ms)
plus an extra coo -> csr materialization. Simple graphs with numeric
weights now build COO arrays straight from the nx adjacency and call
Matrix.from_coo. The node selection preamble mirrors
nx.to_scipy_sparse_array so ordering, subsetting, and error behavior
match exactly; undirected graphs are symmetrized with the same diagonal
correction nx uses (self-loop entries appear as wt + wt - wt under
dup_op=plus, which is exact in IEEE arithmetic).

Kept on the scipy fallback automatically: multigraphs (scipy's
duplicate-coordinate summation matches exactly) and weights that do not
form a 1-D numeric array. That second condition covers non-numeric
attributes, including all-string weights, which infer a <U dtype rather
than object; the fallback raises the identical ValueError. It also
covers array-valued weights: sequence attributes of uniform length
infer a 2-D numeric array, which passes a dtype-kind test but which
from_coo would read as a UDT, so they keep raising ValueError as before
rather than quietly widening what from_networkx accepts.

graphblas without scipy installed can now convert simple graphs
(previously ImportError); multigraphs still need scipy.