fix(indLin): restore the inductive-linearization iteration

[mattfidler]

method="indLin" currently performs one relinearization per hmax-capped
substep. The fixed-point iteration that makes it inductive linearization was
deleted.

Current behavior

Commit bf85a053e ("add indLin method", 2026-06-09) removed case 3: and
case 4: from switch(doIndLin) in src/expm.cpp -- 59 lines. Recover them
with git show bf85a053e^:src/expm.cpp.

Their structure:

• y0 holds the interval-start state and is never updated; w is the iterate.
• Each pass calls meOnly(cSub, w.memptr(), y0.memptr(), ...). meOnly passes
  its yc_ argument to ME as __zzStateVar__ (src/expm.cpp:251), so the
  matrix is built at the latest iterate while propagation always starts from
  y0.
• Loop bound maxsteps = op->mxstep; convergence when
  fabs(w[j]-wLast[j]) < rtol[j]*fabs(w[j]) + atol[j] for every j in
  op->indLin[].

rtol, atol and maxsteps are still declared at src/expm.cpp:329-331 and
are unused in the live body. The current default: branch is
stop(_("unsupported indLin code: %d"), doIndLin), so codes 3/4 error rather
than degrade.

Change

• Reinstate case 3:/case 4: nested inside the hmax substep loop at
  src/expm.cpp:364-396: subdivide [tp,tf] by hmax, iterate to convergence
  within each substep. Pass the state pointer to IndF.
• src/genModelVars.c -- flip fullIndLin to length(wIndLin) > 0.
• src/rxData.cpp -- no change.
• Surface non-convergence. The deleted code returned 1 unconditionally after
  exhausting maxsteps.

Correct value after this issue

A Michaelis-Menten model written with an indLin() forcing converges to the
LSODA answer. Linear models and state-free forcings keep codes 1/2 and are
bit-identical to before.

Test

In tests/testthat/test-ind-lin.R:

• MM via an indLin() forcing matches liblsoda to 1e-5 at the default hmax
• a linear matExp() model is bit-identical to before the change
• hmax still refines the answer at iteration convergence (this pins the order
  of convergence, and is what makes the deferred linear-ramp issue measurable)
• a deliberately non-converging model reports rather than silently returning the
  last iterate

Blocked by #1183 and #1184.