Test the fmm engine guards, string weights, and config key completions; combine repeated indices in csr, csc and GCXS imports - #609

Three behaviors had no test at all in the fast tier. Asking for the
deprecated "fmm"/"fast_matrix_market" Matrix Market engine when
fast_matrix_market is not installed must warn about the deprecation
before it fails, and then fail with an ImportError naming the engine;
neither read nor write had coverage for that, since the existing engine
tests skip themselves when the package is missing. String edge weights
in from_networkx infer a 1-D <U array, so only the dtype-kind half of
the guard sends them down the scipy fallback; the neighboring
array-weight test exercises only the shape half. And the config
mapping's ipython_key_completions hook, which drives config[<tab>]
in IPython, was never called.

Each test was checked by perturbing the line it covers and confirming
it goes red.

Added in the 2026-10-20 coverage round

Behavior fix: repeated indices in an imported csr, csc or GCXS array
(main too). A scipy csr or csc array outside canonical format, and a
pydata GCXS array, may repeat an index within a row (or column), and
from_scipy_sparse imported the arrays unchecked. The Matrix was invalid
(GxB_Matrix_fprint: GrB_INVALID_OBJECT), with nvals counting both entries.

Before: from_pydata_sparse of a 2x3 GCXS with indices [1, 1, 0] in row 0
(values 1, 2, 3) and [2] in row 1 gave nvals 4 with two (0, 1) entries,
and to_dense 2.0 at (0, 1), where pydata's todense gives 3.0.
from_scipy_sparse of the same csr array did the same.

After:

• from_scipy_sparse builds such an array through coo, so repeats need
  dup_op, as they already did in coo input. Without it, a csr or csc with
  a repeated index now raises ValueError "Duplicate indices found" instead
  of building the invalid Matrix.
• from_pydata_sparse sums a GCXS's repeats by default, as pydata's todense
  does (and its COO does when built), with scipy's sum_duplicates on a
  copy, so the caller's arrays are unchanged and any dtype works. dup_op
  combines them otherwise. A 1-d GCXS already summed them (through
  asformat("coo")).
• Unsorted but distinct indices give the same Matrix as before.

New tests: test_from_scipy_sparse_compressed_duplicates (csr, csc) and
test_from_pydata_sparse_gcxs_duplicates (axis 0 and 1; int64, float64,
bool), which fail without the fix.

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}