Restore warning-mode orbit recovery without Newton stalls

[krystophny]

Summary

• restore historical warning-mode continuation for every finite Newton max-iteration iterate while retaining radial damping for bounded VMEC/Boozer/Meiss fields
• recover non-SPECTRE symplectic retry exhaustion with the established adaptive RK/near-axis pusher; keep per-orbit recovery active across a failing axis/edge region or generic cooldown, then reseed and resume the requested symplectic method
• preserve SPECTRE polynomial-extension Newton behavior, harden inverse/sheet paths against non-finite states, and retain meaningful last-valid states
• make Level-1 refraction and forbidden-crossing relocation fall back immediately to the energy-exact Level-0 map when a field sample becomes non-finite
• remove progress, warning, crossing-log, and landing-stat critical/I/O work from the OpenMP tracing region; report aggregate recovery, resume, and intervention counts after the team joins
• preserve large unwrapped toroidal winding while solving the periodic SPECTRE inverse
• stack the independent pyplot module build-order fix from ready PR Order pyplot module before SIMPLE compilation #500 so clean parallel Ninja builds cannot race on pyplot_module.mod

The RK recovery state is local to each orbit. The existing pusher uses its pseudo-Cartesian axis chart below s=0.01; recovery resumes symplectic tracing after s>=0.02. Edge recovery resumes at s<=0.98, while generic failures use exponentially backed-off probes. A successful symplectic step clears the recovery state. Strict mode remains opt-in and unchanged.

This is the ready top PR for the reviewed stack (#494-#498 and #500); the lower PRs remain separated by concern. The branch merges current main and is the exact SIMPLE revision used for the final VMEC/SPECTRE benchmark campaign.

Related: #395, #398, #440, #441, #442, #469, #493, #317. Runtime builds pin libneo 79dc370 = current main e0437ad (near-axis continuation and full-torus chartmap inverse) plus exactly the reviewed NVFORTRAN portability change from open libneo PR #394.

Validation

• combined fo at exact head c82d155: static analysis, build, complete test suite, and lint passed in 381.0 s
• focused recovery-region, Newton-status, physical-exit, and SPECTRE-crossing tests pass
• isolated PR Order pyplot module before SIMPLE compilation #500 clean fo build and full fo passed
• GDB confirmed the prior VMEC slowdown was repeated nine-level retry descent after each one-step RK fallback, not angle divergence, idle OpenMP workers, or a deadlock
• exact NVHPC build and expanded cluster smoke (all five algorithms plus the exact 256-marker VMEC production population) follow in the benchmark repository