Remove TorchScript helpers from linear CG

[saitcakmak]

Summary

• remove the deprecated torch.jit.script helpers from linear_cg
• consolidate preconditioned and unpreconditioned iterations into one code path
• add direct coverage for a custom Jacobi preconditioner

Why

The helpers were introduced as TorchScript-only optimization boundaries. Replacing them with torch.compile, as proposed in #129, adds substantial first-call compilation overhead and shape-specialization costs for small fragments inside a Python loop. The default preconditioner already implements the unpreconditioned behavior by cloning the residual, so both branches can share the same update code without either compilation mechanism.

This also removes the torch.jit.script deprecation warning emitted when importing linear_operator on recent PyTorch versions.

Correctness

The consolidated loop preserves the existing CG update order. A new float64 parity test compares a Jacobi-preconditioned solve with torch.linalg.solve and verifies that the supplied preconditioner is used repeatedly. Temporarily reversing the residual-update sign caused this test to fail, confirming that it detects an incorrect update.

Test plan

• ufmt on both changed files
• python -m flake8 -j 1 linear_operator/utils/linear_cg.py test/utils/test_linear_cg.py
• python -m pytest test/utils/test_linear_cg.py test/functions/test_inv_quad_logdet.py test/operators/test_added_diag_linear_operator.py test/operators/test_low_rank_root_added_diag_linear_operator.py -xvs (403 passed)
• python -Wd -c "import linear_operator" (no TorchScript deprecation warning)

[kayween]

It seems good to me to remove torch.jit.script given that it's already deprecated a long time ago.