Carry B_s to the wout file on the full radial mesh

[CharlesCNorton]

Fixes #22.

jxbforce builds B_s on the full radial mesh in a local array and drops it at RETURN, so wrout receives the half-grid array that bss produced, applies a full-mesh endpoint rule to it and writes it as !Full mesh. bsubsh becomes INTENT(inout) and carries the full-mesh values back, which is what PARVMEC jxbforce.f:33 does. The note at wrout.f90:603 that recorded the mismatch goes with it.

Compute_Currents then has to read bsubsmn on the full mesh rather than averaging two half-grid surfaces into t1. ORNL-Fusion/LIBSTELL read_wout_mod.f90, the copy PARVMEC links against, makes the same four substitutions.

Checked against the PARVMEC wout files for input.solovev, input.cma and input.cth_like_fixed_bdy, as the largest difference over the array relative to its maximum:

	bsubsmns	currumnc	currvmnc
solovev	8.0e-3 -> 9.4e-6	1.5e-10 -> 1.5e-10	4.2e-4 -> 2.5e-11
cma	2.5e-1 -> 5.1e-3	1.0e+0 -> 1.5e-2	9.8e-1 -> 2.6e-2
cth_like_fixed_bdy	2.5e-1 -> 2.2e-4	8.9e-1 -> 2.1e-6	5.1e-1 -> 2.5e-6

The boundary value is the second commit. PARVMEC reads bsubs(ns,:) on the right-hand side of its own edge rule, which is the half-grid value still sitting in the array it shares with the caller; here that value is bsubsh(ns,:), so the outermost half point can be used directly instead of extrapolating from two full-mesh points. The largest disagreement with PARVMEC on the boundary surface goes from 9.4e-6 to 7.2e-13 for solovev and from 1.5e-3 to 7.0e-7 for cth_like_fixed_bdy, level with every interior surface. Compute_Currents runs from 2 to ns-1 and extrapolates the ends, so it never reads that row and the currents are unchanged. The cma residuals are larger because that pair of runs converged to slightly different states: niter differs by 7 percent and bmnc by 3e-4 independently of this change.

lbsubs = T now reaches the wout, since the correction is applied to the array the caller keeps.

VMEC++ writes the half-grid values under the name bsubsmns, to match this code, and publishes the full-grid transform separately as bsubsmns_full; the note at output_quantities.cc:5159 records the mismatch with the wout file's own declaration and leaves it open. After this change educational_VMEC and PARVMEC agree on that variable.