Keep the istft slice from emptying when fft_length is 0 or 1

[MarcosAsh]

Description

istft returned an empty array whenever fft_length was 0 or 1 with center=True. The end of the slice was -(fft_length // 2), which is -0 in that case, so x[..., start:0] kept nothing. jax counts back from the expected length instead, expected_output_len - (fft_length // 2), which this now matches.

The test_istft7 and test_istft8 exclusions stay. Removing them passes on linux but fails on macOS, where the same input comes back as the right values interleaved with zeros, [v0, 0, v1, 0, ...] against [v0, v1, v2, ...]. That is a separate Metal only problem in the overlap and add path and it needs a machine with Metal to chase, so it is not in this PR.

The other five istft exclusions also stay. Their mismatches are all inside the 5% of each boundary that math_test.py already truncates for numpy, jax, torch and openvino, so they are the same boundary difference those backends have rather than anything wrong here. Adding mlx to that list is a keras change, not one for this repo.