`SegmentedPolynomial` checkpoints saved on CUDA cannot be loaded on CPU due to missing `__reduce__` on backend classes, preventing model portability

[jchodera]

(prepared with the help of claude code)

Describe the bug
SegmentedPolynomial checkpoints saved on CUDA cannot be loaded on CPU due to missing __reduce__ on backend classes, preventing model portability.

SegmentedPolynomial.__init__ selects a backend at construction time based on whether cuequivariance_ops_torch is installed:
 - CUDA machine → self.m = SegmentedPolynomialFromUniform1dJit(...)
 - CPU machine → self.m = SegmentedPolynomialNaive(...)
PyTorch's default nn.Module pickling freezes the class type into the checkpoint. When the checkpoint is loaded on a CPU-only machine, SegmentedPolynomialFromUniform1dJit is reconstructed directly (bypassing __init__), but tensor_product_uniform_1d_jit is None at the module level. The first forward() call raises:
TypeError: 'NoneType' object is not callable

Fix: Add __reduce__ to SegmentedPolynomialFromUniform1dJit (and FusedTP, IndexedLinear) to re-delegate to SegmentedPolynomial.__init__ at unpickle time, which re-runs the HAS_CUE_OPS check and selects the correct backend for the loading machine. Requires storing the original polynomial and math_dtype arguments as instance attributes.

 Affected files:
 - cuequivariance_torch/primitives/segmented_polynomial_uniform_1d.py — SegmentedPolynomialFromUniform1dJit (lines 89–301)
 - cuequivariance_torch/primitives/segmented_polynomial_fused_tp.py — SegmentedPolynomialFusedTP
 - cuequivariance_torch/primitives/segmented_polynomial_indexed_linear.py — SegmentedPolynomialIndexedLinear
 - cuequivariance_torch/primitives/segmented_polynomial.py — SegmentedPolynomial (line 178–203 for the backend selection, no __reduce__)

To Reproduce
This Python script illustrates the issue (requires torch, cuequivariance, cuequivariance_torch):
issue-1-cuequivariance-checkpoint-portability-mre.py

Expected behavior
Serializing/deserializing models based on cuequivariance should use the hardware available on the current system, rather than the system used when the model was initially constructed.

GPU HW/SW(please complete the following information):
Environment:
torch version: 2.11.0+cu130
cuequivariance: 0.9.1
cuequivariance_torch: 0.9.1