mcda GPU paths fail: piecewise standardize on cupy, piecewise/categorical on dask+cupy, monte_carlo sensitivity on cupy

[brendancol]

Found by the performance sweep against xrspatial/mcda, verified on a CUDA host (cupy 13.6). standardize() documents support for numpy, cupy, dask+numpy, and dask+cupy, but three GPU paths fail outright.

1. standardize(method="piecewise") on cupy input

_piecewise calls xp.interp(data, bp, vl) with xp = cupy while bp and vl are numpy arrays:

NotImplementedError: Only int or ndarray are supported for a

2. standardize(method="piecewise" | "categorical") on dask+cupy input

The map_blocks chunk functions call np.asarray(block) with a comment claiming this "handles cupy chunks". It does not; cupy refuses implicit conversion:

TypeError: Implicit conversion to a NumPy array is not allowed. Please use `.get()` to construct a NumPy array explicitly.

Even if the conversion worked, it would be a device-to-host copy per chunk with the interpolation running on the CPU and the result left there.

3. sensitivity(method="monte_carlo") on cupy input

_monte_carlo reads score.values every iteration, which hits the same implicit-conversion TypeError on cupy-backed data. Were the conversion allowed, it would be one device-to-host transfer per sample (1000 by default) with the Welford accumulation running on the host.

Proposed fixes

• piecewise on cupy: hand cupy.interp cupy copies of the breakpoint and value tables so everything stays on device.
• piecewise/categorical chunk functions: detect cupy blocks and use the matching array module instead of converting to numpy.
• monte_carlo: accumulate with the array's own module (score.data rather than .values) and convert to numpy once at the end.