Drop AdditiveSchwarzDiagnostics from public API

[schroedk]

Summary

• Removes AdditiveSchwarzDiagnostics (Rust struct, pub use re-exports, Python class, additive_schwarz_diagnostics accessors at every layer).
• Inlines the 5 scheduling metrics directly into AdditiveScheduler — they remain private to the Auto heuristic.
• Keeps ReductionStrategy and the Preconditioner::Additive(_, ReductionStrategy) knob unchanged: the wall-clock-vs-memory tradeoff (12x peak memory at scale) is a real user concern.

Closes #34.

Why

Empirical bake-off across small / medium / huge regimes (see the issue comments) showed:

• Auto's wall-clock decisions are within the ~5% measurement noise floor on every regime in the project's normal operating range.
• AdditiveSchwarzDiagnostics had zero readers outside the scheduler itself — no test asserted on its fields, no benchmark inspected it, no production caller read it.
• The heuristic's tuning constants are private, so an external caller could not reproduce or override decisions from the diagnostics anyway.

ReductionStrategy is kept because memory matters: on a 100K-DOF problem, ParallelReduction peaks at ~9.6 MB vs AtomicScatter's ~800 KB (12x). At 1M DOFs that becomes ~96 MB vs ~8 MB. Memory-constrained or server-scenario users need the override.

Test plan

• cargo build --workspace --release clean
• cargo test --workspace --exclude within-py all pass (within-py lib-test failure is the expected "PyO3 can't run standalone" issue per CLAUDE.md)
• pixi run develop rebuilds the extension cleanly
• pytest tests/ 97/97 pass
• Within nested-rayon regression test rewritten to use subdomains().len() and direct inner_parallelism_work_estimate() iteration — same assertions, no diagnostics() dependency