focal: missing dask+cupy backends force `.compute()`

[brendancol]

Summary

focal.mean(), focal.apply(), and focal_stats() raise NotImplementedError for dask+cupy inputs. Users with GPU clusters are forced to .compute() first, negating dask's memory benefits and causing OOM on large datasets.

Affected Functions

Line	Function	Issue
focal.py:164	mean()	NotImplementedError for dask+cupy
focal.py:502	apply()	NotImplementedError for dask+cupy
focal.py:922	focal_stats()	NotImplementedError for dask+cupy

Severity

MODERATE — only affects dask+cupy users, but forces full materialisation which causes OOM on large datasets.

Suggested Fix

• Implement dask+cupy backends using map_overlap with cupy kernels (similar to the existing dask+numpy path)
• The pattern is well-established in other spatial ops in this codebase (e.g., slope, aspect, convolution)