fix(image): guard spectral_angle_mapper against NaN on zero-norm pixels

[xodn348]

Summary

spectral_angle_mapper (and the SpectralAngleMapper class) silently returns NaN when any
spatial pixel has an all-zero channel vector — for example a constant-black pixel in an image.
The SAM score is computed as acos(dot / (|preds| * |target|)), and when either norm is zero the
quotient is 0/0 = NaN, which then propagates through torch.clamp and acos into the final
reduction. The fix clamps preds_norm * target_norm to torch.finfo(dtype).eps before dividing,
matching the convention used by torch.nn.functional.cosine_similarity. A zero-norm pixel now
contributes acos(0) = π/2 (90°) to the mean score instead of poisoning it with NaN.

Issue

Fixes #3322

Local verification

$ cd /tmp/torchmetrics
$ python -m pytest tests/unittests/image/test_sam.py -v -k "not ddp and not half"
============================= test session starts ==============================
platform linux -- Python 3.11.15, pytest-9.0.3, pluggy-1.6.0 -- /usr/local/bin/python
cachedir: .pytest_cache
rootdir: /tmp/torchmetrics
configfile: pyproject.toml
plugins: xdist-3.8.0, doctestplus-1.7.1
collected 45 items / 24 deselected / 21 selected

tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds0-target0-sum] PASSED [  4%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds0-target0-elementwise_mean] PASSED [  9%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds1-target1-sum] PASSED [ 14%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds1-target1-elementwise_mean] PASSED [ 19%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds2-target2-sum] PASSED [ 23%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds2-target2-elementwise_mean] PASSED [ 28%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds3-target3-sum] PASSED [ 33%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam[False-preds3-target3-elementwise_mean] PASSED [ 38%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds0-target0-sum] PASSED [ 42%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds0-target0-elementwise_mean] PASSED [ 47%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds1-target1-sum] PASSED [ 52%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds1-target1-elementwise_mean] PASSED [ 57%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds2-target2-sum] PASSED [ 61%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds2-target2-elementwise_mean] PASSED [ 66%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds3-target3-sum] PASSED [ 71%]
tests/unittests/image/test_sam.py::TestSpectralAngleMapper::test_sam_functional[preds3-target3-elementwise_mean] PASSED [ 76%]
tests/unittests/image/test_sam.py::test_error_on_different_shape PASSED        [ 80%]
tests/unittests/image/test_sam.py::test_error_on_invalid_shape PASSED          [ 85%]
tests/unittests/image/test_sam.py::test_error_on_invalid_type PASSED           [ 90%]
tests/unittests/image/test_sam.py::test_error_on_grayscale_image PASSED        [ 95%]
tests/unittests/image/test_sam.py::test_sam_no_nan_on_zero_pixel PASSED        [100%]

================ 21 passed, 24 deselected, 48 warnings in 3.41s ================
=== LOCAL_TEST_PASSED ===

Risk

The only changed line is in _sam_compute; existing random-valued inputs never produce a
zero-norm pixel so all current tests pass unchanged. The behaviour for non-degenerate inputs
is numerically identical because preds_norm * target_norm ≥ eps already holds. Any caller
that was deliberately testing for NaN on zero-input would observe π/2 instead, which is a
breaking change only in that narrow and arguably-incorrect use case.

---

📚 Documentation preview 📚: https://torchmetrics--3372.org.readthedocs.build/en/3372/