feat(indLin): route nonlinear terms into indLin() when converting an ODE model

[mattfidler]

Converting a d/dt() model puts nonlinear terms inside the rate constants,
where they cannot be iterated and where they violate an assumption the
event-sensitivity code already documents.

Updated while implementing this in #1213. Two things in the original
write-up turned out to be wrong and are corrected below: consuming the flag
column is not sufficient (it is per-equation, not per-term), and
rxIndLinStrategy()/rxIndLinState() cannot keep working the way this issue
assumed. A hand-written state-dependent k_from_to is also now a parse error
rather than being left alone -- a deliberate change of scope, not a
discovery.

Current behavior

• indLin() (R/indLin.R:9-109) calls rxIndLin_() (:25), which returns a
  character matrix carrying the state columns, a _rxF column and an indLin
  flag column (built at src/expm.cpp:37-48). Lines :28-29 keep only the state
  block and _rxF, and discard the flag column.
• So a Michaelis-Menten model converts to
  k_central_output = vmax/(v*(km + central/v)) -- a state inside a rate
  constant. Verified against current HEAD.
• That contradicts what the event-sensitivity jump code states at
  inst/include/rxode2parseHandleEvid.h:1710-1719: "indLin's whole premise is a
  CONSTANT (time-invariant) Jacobian over the interval -- the k_from_to rate
  constants are parameter-functions only, never state-dependent".
• It is also why rxSolve(<mm ode model>, method="indLin") is ~70% off against
  liblsoda at the default hmax, measured on the 13-point sampling grid from
  Sharif et al. 2022.

Change

R/indLin.R:9-109 -- consume the flag column. Corrected: the flag is
per equation, not per term -- .rxIndLinLine() returns one .fullIndLin
for the whole d/dt(state) line (R/rxIndLin.R:167) -- so consuming it cannot
produce the target shape below, which keeps ka*depot linear while routing only
the Michaelis-Menten residual to the forcing. Term-level routing is needed, and
the place to do it is where the per-term decision is already made and where the
flag is currently set: the three branches of .rxIndLinLine() that set
.fullIndLin <- TRUE (R/rxIndLin.R:95, :100, :115) route the whole term
to _rxF instead of emitting a state-dependent coefficient into A.

rxIndLinStrategy() / rxIndLinState() (R/rxIndLin.R:19,30) keep
controlling how a multi-state product is split. Corrected: they cannot.
Dividing any one state out of a multi-state product always leaves a coefficient
that still reads a state -- van der Pol's mu*y^2*dy leaves mu*y*dy or
mu*y^2 whichever is chosen -- so such a term goes to the forcing regardless of
the preference. Both stay exported and documented, with an @details note that
they no longer affect the split.

Target shape for the Michaelis-Menten example (unchanged, and achieved):

matExp()
cmt(depot)
cmt(central)
k_depot_central <- ka
cp <- central/v
indLin(central) <- -vmax*cp/(km + cp)

Correct value after this issue

rxSolve(<ode model>, method="indLin") is accurate at the default hmax.

Hand-written state-dependent k_from_to models are untouched and keep
working. Superseded: a matExp() rate constant that depends on a
compartment is now a parse error, naming the constant, the compartment it
reaches and the indLin() form to use instead. The matrix exponential is only
valid when the rate matrix is constant over the step, so such a model was
silently wrong rather than merely unsupported. Models generated by
rxSensMatExp() are exempt for now: their sensitivity blocks are built out of
rate constants throughout and reference the primal ones by name, so rewriting
that generator is separate work and is not done here.

Test

• indLin(mmOde) emits indLin(central) <- ... with no k_central_output at
  all (the elimination is entirely in the forcing)
• solving matches liblsoda at the default settings -- measured relative error
  3.9e-06 against the ~70% above
• the van der Pol block that exercises rxIndLinStrategy()/rxIndLinState()
  still runs unchanged Corrected: that block was rewritten rather than
  re-enabled. Its method="lsoda" arm was never lsoda -- rxSolve.default
  force-selects method="indLin" for any model carrying an indLin descriptor,
  so it was comparing the model against itself and the assertion was vacuous. It
  now compares against a plain ODE model, which is what the commented-out
  assertions were reaching for, and additionally asserts that the three strategy
  variants produce identical models.

Resolution

Implemented in #1213 (branch fix/indlin-iteration), together with #1185.