Fix Upsample align_corners singleton output

[ishtihoss]

Summary

Fix nn.Upsample with align_corners=True when an output spatial dimension is 1.

The aligned-corners scale used (N - 1) / (M - 1), so M == 1 raised a ZeroDivisionError. Using max(M - 1, 1) keeps existing behavior for M > 1 and maps singleton outputs to the first input coordinate.

Tests

• black --check python/mlx/nn/layers/upsample.py python/tests/test_nn.py
• python -m unittest test_nn.TestLayers.test_upsample_align_corners_one_dim -v
• python -m unittest test_nn.TestLayers.test_upsample -v

[zcbenz]

Thanks!