Fix RK near-axis pseudo-Cartesian coordinates

[krystophny]

Summary

Fix the RK near-axis pseudo-Cartesian transform used by orbit_timestep_axis.

The old path entered the near-axis chart as (s cos theta, s sin theta) but treated it as (sqrt(s) cos theta, sqrt(s) sin theta) when evaluating velocity and returning to flux coordinates. That made the axis an artificial sink for some mirror-trapped particles.

This PR makes the chart consistent:

• enter with (sqrt(s) cos theta, sqrt(s) sin theta)
• recover s = x^2 + y^2
• use the matching 0.5 chain-rule factor in velo_axis
• keep the floor only for field evaluation, not as the physical radial mapping

This is the RK-only extraction from PR #401. The symplectic near-axis experiment stays stacked there.

Verification

fo check

{"ok":true,"stage":"done","target":"","summary":"build and tests passed","hint":"","rerun":"","log_path":"","elapsed_s":2.401,"modules":469,"cached":0,"changed":469}