convolve_2d truncates float64 input to float32

[brendancol]

Describe the bug

All four convolve_2d backends hardcode .astype(np.float32) on input data and np.float32/f4 on output arrays. When float64 data is passed, half the mantissa bits are silently discarded.

Affected functions:

• _convolve_2d_numpy (line 290): data = data.astype(np.float32)
• _convolve_2d_dask_numpy (line 332): data = data.astype(np.float32)
• _convolve_2d_cupy (line 396): data = data.astype(cupy.float32)
• _convolve_2d_dask_cupy (line 405): data = data.astype(cupy.float32)

convolve_2d is used internally by convolution_2d, focal.mean, focal.hotspots, and other downstream functions.

Expected behavior

Integer inputs should be promoted to float32 (to avoid integer overflow in weighted sums). Float64 inputs should stay float64. Output dtype should match the promoted input dtype.

Reproduction

import numpy as np
from xrspatial.convolution import convolve_2d

data = np.array([[1e7 + 0.01, 1e7 + 0.02],
                 [1e7 + 0.03, 1e7 + 0.04],
                 [1e7 + 0.05, 1e7 + 0.06],
                 [1e7 + 0.07, 1e7 + 0.08]], dtype=np.float64)
kernel = np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.float64)

result = convolve_2d(data, kernel)
print(result.dtype)   # float32 -- should be float64
# The 0.0x differences are lost in float32

Affected code

xrspatial/convolution.py -- all four backend functions for convolve_2d