Test coverage: morphology edge inputs (Inf, all-NaN, strip, integer)

[brendancol]

Closes #3404

Adds test-only coverage for four morphology edge inputs that the source already handles correctly but no test exercised. No source changes.

• Inf / -Inf inputs: confirm they flow through the min/max kernel and the NaN-propagation branch without turning into NaN
• All-NaN raster: confirm every op returns all NaN
• Nx1 / 1xN strip rasters: confirm shape is preserved on the dask map_overlap path where one axis is narrower than the overlap depth
• Integer-dtype input: confirm it is accepted and promoted to float64

Backends: numpy plus dask, cupy, and dask+cupy parity where applicable, using the cross-backend helpers in general_checks.py. The cupy and dask+cupy variants ran on a GPU host before this PR.

Test plan

• new tests pass on numpy, dask, cupy, dask+cupy (GPU host)
• full test_morphology.py and test_morphology_derived.py pass (86 tests)

[brendancol]

PR Review: morphology edge-input coverage (Inf, all-NaN, strip, integer)

Test-only PR. No source files changed, so the correctness / backend / performance / docs categories are mostly N/A. I reviewed the four new test groups in test_morphology.py.

Blockers

None.

Suggestions

None blocking.

Nits

• The all-NaN, strip, and integer-dtype tests cover numpy and dask but not cupy. That is fine here: the GPU min/max kernel logic is identical and is already cross-checked by the broad erode/dilate cupy parity tests plus the new Inf cupy parity test. Adding cupy variants would be belt-and-suspenders, not a real gap.

What looks good

• Inf tests assert concrete min/max values and finite-vs-inf positions, not just "runs without error".
• Inf parity covered on numpy, dask, and cupy.
• Strip and integer tests are parametrized over both orientations and three integer dtypes.
• all-NaN uses boundary='nearest', so the assertion isolates the all-NaN input rather than edge padding.
• New tests pass alongside the existing 86-test suite.

Checklist

• Source unchanged, no algorithm to verify
• Backends consistent: parity asserted for Inf (numpy/dask/cupy), all-NaN and strip (numpy/dask)
• NaN handling correct
• Edge cases covered (the point of the PR)
• Dask boundaries: strip tests hit the degenerate-axis map_overlap path
• No premature materialization: .compute() only in assertions
• Benchmark not needed, test-only
• README matrix not needed, no new function
• Docstrings present on each test