spy() bound the ``fig`` local only when it created the figure itself, so
``spy(A, figure=fig)`` crashed creating the axes and ``spy(A, axes=ax)``
crashed in the auto-markersize path (found while writing the smoke
tests). The figure kwarg is now used when given, and auto-markersize
reads dpi from ``axes.figure``, which is correct in all three call
forms. Regression tests cover both previously-crashing forms.

…iz.draw

Two bugs reported in the pyOpenSci review.  Both change viz.py, so they
are itemized in a single commit.

gh-473: datashade binned element (r, c) into the pixel spanning
[c, c+1) x [r, r+1), so it rendered centered at (c+0.5, r+0.5), half a
cell off the tick labeled (c, r); spy centers the same element exactly
on the tick.  The axis limits now use the imshow integer-center
convention (-0.5 to n-0.5), making datashade agree with spy.  The
issue's other symptom (elements invisible until zoom-out) does not
reproduce on the current holoviews/hvplot/bokeh stack; it was a
2023-era library artifact.

gh-474: reciprocal directed edges drew as coincident straight lines
with both weight labels on the same midpoint, hiding one weight.
Reciprocal pairs now draw with an arc (connectionstyle arc3, rad 0.1)
and matching label placement, so both arrows and both weights are
visible; other edges and self-loops stay straight.

The gh-474 fix needs networkx 3.3, the release that gave
draw_networkx_edge_labels its connectionstyle parameter.  We support
networkx >=2.8, so draw() feature-detects that parameter and keeps the
previous straight rendering when it is missing; passing it to an older
networkx raises TypeError.  Curving the edges but not the labels would
be worse than not curving at all, since the labels would sit back on the
shared chord midpoint (the overlap gh-474 is about) and would no longer
track their arrows.

Both fixes carry display-free regression tests: datashade pixel centers
match spy's convention, and draw places the two weight labels at
separated anchors.  The label check measures separation as a fraction of
the edge length rather than comparing positions for exact inequality,
because networkx returns two midpoint anchors that differ by
floating-point noise even when the labels coincide on screen.  A third
test stands in a pre-3.3 signature to cover the fallback.

Removes the coverage omit block (viz.py was its only entry, with a TODO
to un-omit once tests existed; they do now).  The viz test module covers
87.6% of viz.py counting statements and branches, 91.1% counting
statements alone.  The misses are the optional-import failure path, the
interactive plt.show() path, and guards for inputs the tests do not
construct: matrices too large for int64 indices, ragged aggregator
grids, and caller-supplied opts_kwargs.