[2.6/2.7] Rust SOR + multigrid Laplace kernels

[RFingAdam]

Acceptance criteria

• crates/atlc3_kernel/src/laplace/sor.rs implements 5-pt FD with rayon parallelism
  (stub only)
• PyO3 binding atlc3._kernel.laplace_sor(...)
• 10× faster than NumPy on 1000×1000 microstrip
• Multigrid V-cycle ≥5× faster than SOR on 2000×2000

Status: ❌ Not started — moved to Phase 5

The Python+NumPy+PyAMG path covers practical PCB cross-sections (a few
thousand pixels). Rust acceleration is an optimization, not a correctness gap.

[RFingAdam]

Moved to Phase 5 milestone. NumPy + PyAMG + scipy.sparse remains the production solver path; Rust acceleration is an optimization, not a correctness gap.

Inline rationale added to the stub files:

• crates/lineforge_kernel/src/laplace/sor.rs
• crates/lineforge_kernel/src/laplace/multigrid.rs

Both now note Phase 5 + the production path so contributors don't think the kernel is missing.