BUG: preserve negative FFT interpolation weights - #20482

Description

Fixes #20403.

convolve_fft() could return 0.0 instead of an interpolated value when all of the following were true:

• nan_treatment="interpolate" was used,
• the input contained a NaN, and
• the normalized kernel contained negative coefficients.

Before

The FFT interpolation path computes a per-pixel normalization weight. The code treated every weight smaller than a tiny positive threshold as “no weight” and replaced that output with zero.

That condition also matched valid negative weights produced by kernels with negative coefficients. As a result, valid interpolation results were discarded.

For the reproduction in #20403:

• convolve(...) returned 26.666666666666668
• convolve_fft(...) returned 0.0

After

The code now treats only weights whose absolute value is close to zero as absent. Valid negative weights are retained and used for normalization.

With the same input:

• convolve(...) returns 26.666666666666668
• convolve_fft(...) returns 26.666666666666664

The tiny difference is normal floating-point FFT precision.

Code changes

• Updated the no-weight check in convolve_fft() to use the absolute normalization weight.
• Added a regression test using a NaN-containing image and a normalized kernel with negative coefficients.
• Added a convolution changelog entry.

Verification

I reproduced the issue locally against the latest main branch before making the change.

Tests run:

• New regression test for negative interpolation weights
• Full astropy/convolution/tests/test_convolve_fft.py suite: 620 passed

AI Disclosure

I used Codex (GPT-5) to help inspect the repository, explain the contributing guidance, assist with the initial code and test draft, and format this pull request description.

I reproduced the issue locally, reviewed the final code and tests, ran the verification commands, and take full responsibility for this pull request and all reviewer interactions.

• I certify that I am human and that I take full responsibility for this pull request including all interactions with reviewers.

Merge method

• By checking this box, the PR author has requested that maintainers do NOT use the "Squash and Merge" button. Maintainers should respect this when possible; however, the final decision is at the discretion of the maintainer that merges the PR.