Style cleanup in focal.py: unused import, isort, mutable default (#2731)

[brendancol]

Closes #2731

Style-only cleanup of xrspatial/focal.py, flagged by the project's own setup.cfg (flake8, isort) and a mutable-default grep. From the deep-sweep style pass on the focal module.

• Cat 3 (F401): removed the unused not_implemented_func import. No __all__, nothing imports it from xrspatial.focal, so it wasn't a re-export.
• Cat 4 (isort): reordered and rewrapped the top-of-file import block to line_length=100. Mechanical, no functional change.
• Cat 5 (mutable default): mean(..., excludes=[np.nan], ...) becomes excludes=None, resolved to [np.nan] in the body. The list was never mutated, so behaviour is identical. Added test_mean_default_excludes_does_not_leak.

No behavioural change is intended for the Cat 3/4 fixes. The Cat 5 fix is behaviour-preserving. No config widening, no autoformatter.

Backends: static source-level changes, uniform across numpy / cupy / dask+numpy / dask+cupy. mean dispatches through ArrayTypeFunctionMapping; no path-specific logic touched.

Test plan:

• flake8 xrspatial/focal.py clean
• isort --check-only xrspatial/focal.py clean
• 115 focal tests pass (incl. GPU/dask mean tests, CUDA available)
• new regression test passes

[brendancol]

Domain-aware review (deep-sweep style worker)

Scope check: the diff is style-only and matches the three findings in #2731.

Correctness:

• The excludes is None -> [np.nan] resolution sits at the top of mean(), ahead of both the 3D _apply_per_band branch and the 2D loop, so both paths get the intended default. Behaviour is unchanged from the old excludes=[np.nan] default.
• isort output wraps the xrspatial.convolution and xrspatial.utils from-imports onto continuation lines at the 100-char limit; flake8 is clean on the result (no E128/E501), so the wrapping is consistent with the project config.
• not_implemented_func removal is safe: no __all__, and nothing imports it from xrspatial.focal.

Backends: changes are static source edits and apply uniformly across numpy / cupy / dask+numpy / dask+cupy. mean dispatches through ArrayTypeFunctionMapping; no per-backend path was touched. The full focal suite (115 tests, including the GPU and dask mean tests) passes with CUDA available.

Nit (non-blocking): test_mean_default_excludes_does_not_leak runs on data_random, which has no NaNs, so it proves None == [np.nan] equivalence and no cross-call leak but does not itself exercise the NaN-exclusion path. That path is already covered by the existing test_mean_* transfer-function tests, so I left it as is.

No blockers, no required changes.