Phase 4 — parallel Jacobi sweeps across NeuronCores

[scttfrdmn]

Roadmap phase tracker

This issue tracks trnsolver's work on Phase 4 of the trnsci roadmap.

See the suite-level roadmap
for the full phase matrix and cross-project dependencies. A reader-friendly
version lives at trnsci.dev/roadmap/.

What this phase means for trnsolver:
Jacobi rotations on disjoint pairs commute, so a sweep is
embarrassingly parallel. Phase 4 exploits that for large-n eigh.

Done means:

• Multi-chip Jacobi: disjoint rotation pairs assigned per NeuronCore, sweep convergence check via all-reduce.
• Large SCF eigh (n > 4096) runs on trn1.32xlarge within memory.
• Benchmarks at trnsci.dev/trnsolver/benchmarks/.
• Parallel CG / GMRES with sharded preconditioners (depends on Phase 3).

Coordination

• Label: phase-4-multichip — matches the same label in every sub-project.
• Cross-project dependencies, if any, are called out in the "Done means"
  list above. Link any PRs / child issues here for tracking.

Close this issue when all "Done means" items are satisfied.