`SphericalHarmonics` naive backend does not support `torch.compile`

Describe the bug
The descriptors.spherical_harmonics API causes torch.compile to fail. This is critical because it is automatically used as a fallback in certain cases, meaning it is strictly broken on the CI where cuda is unavailable

To Reproduce

import torch
import cuequivariance as cue
import cuequivariance_torch as cuet
from cuequivariance import descriptors

e = descriptors.spherical_harmonics(cue.SO3(1), [0, 1, 2])
f = cuet.SegmentedPolynomial(e.polynomial, method="naive").to("cpu")
x = torch.randn(4, 3)

# Fails with:
#   Unsupported method call: Dynamo does not know how to trace
#   method `__add__` of class `Subscripts`.
torch.compile(f, fullgraph=True)([x])

Expected behavior
The spherical_harmonics method should use a custom_op to be compatible with torch.compile.

GPU HW/SW(please complete the following information):

• CUDA toolkit version: Repro does not require CUDA
• torch or ngc docker version: 2.8
• Driver version: Repro does not require CUDA
• full name of GPU: Repro does not require CUDA

Start with descriptors.spherical_harmonics and cuet.SegmentedPolynomial(..., method="naive"), then run the provided CPU torch.compile reproduction with PyTorch 2.8. Trace where the naive spherical harmonics path invokes Subscripts and review the custom_op integration there. Done means the reproduction succeeds with fullgraph=True without requiring CUDA.