Track vmecpp tokamak split and environment cleanup

[krystophny]

Goal

Bring the tokamak-support work back under control, keep a clean path from upstream/main, and split the current large branch into reviewable PRs with clear verification at each step.

Current situation

• Upstream-facing tokamak work currently lives in PR Tokamak rebuild proximafusion/vmecpp#478 and is too broad.
• That branch currently mixes tokamak solver work with Python 3.14 packaging work.
• Python 3.14 is still blocked on the simsopt fix at Enable Python 3.14 source builds hiddenSymmetries/simsopt#619.
• Local pre-commit under Python 3.14 also runs into a hook dependency problem (docformatter -> untokenize), which is another reason to separate environment work from 3.14 support.

Checklist

• Review every commit on tokamak-rebuild against upstream/main and classify it as keep, split, or drop.
• Write down the minimal PR split that makes technical sense.
• Update PR Tokamak rebuild proximafusion/vmecpp#478 to state clearly that it is a temporary umbrella PR for tokamak support and will be split.
• Keep the Python 3.14 support work separate and link it to the simsopt blocker: Enable Python 3.14 source builds hiddenSymmetries/simsopt#619.

Environment first

• Create a fresh branch from upstream/main for a Nix flake.
• Mirror the supported CI environment as closely as possible, starting from the Ubuntu 22.04 + Python 3.10 test job and the conda environment.
• Record which compiler, Python, CMake, NetCDF, HDF5, LAPACK, OpenMP, and MPI versions the flake provides.
• Verify that upstream/main builds and tests in that flake without the Python 3.14 version bump.
• File a small PR for the Nix flake only.

Python 3.14 support

• Keep the Python 3.14 branch scoped to packaging and compatibility only.
• Re-run install and focused tests once simsopt support is available upstream.
• Remove temporary fork references once upstream dependencies support 3.14.

Tokamak support split

• Rebuild the tokamak work from the clean base branch instead of continuing to pile changes onto the old exploratory line.
• Split out asymmetric input and INDATA fixes.
• Split out asymmetric inverse DFT support.
• Split out magnetic-axis recovery fixes.
• Split out force-symmetry / force-DFT fixes.
• Split out the multithreaded asymmetric DFT indexing fix.
• Split out any output / wout handling changes only if they are independently justified.

Verification and comparisons

• Keep circular_tokamak as the symmetric control case.
• Keep up_down_asymmetric_tokamak as the primary target case.
• Compare against educational_VMEC and VMEC2000 at every physics-sensitive step.
• Use jVMEC as an independent cross-check, especially for coefficient conventions and convergence behavior, but not as the sole scalar reference.
• Add or keep focused regression tests for every accepted fix.
• Keep benchmark reports and analysis notes, but tighten them so they document signal instead of churn.

Notes

Use this issue as the anchor for branch descriptions, PR descriptions, and follow-up commits while the split is in progress.