fix: linCmtB(which1 = -3) now handles infusions (#1236)

[mattfidler]

Summary

Fixes #1236. linCmtB(which1 = -3) -- the dose-time (moving-boundary)
sensitivity added in #1235 -- previously returned NA_REAL for any individual
whose regimen contained an infusion, because the derivative needs dA/dt
(which includes the infusion rate), and ind->InfusionRate is only maintained
while solving. rxode2_df.cpp's output pass recomputes lhs from the saved
amounts after iniSubject() clears the rates, so the rate read there was
always 0.

• linCmtB() now caches the live infusion rate per output index while it is
  genuinely solving that index, mirroring the existing fx/Jacobian Alast
  cache (getAdvan(idx)/lc.restoreFx()/lc.restoreJac()), keyed on the same
  !ind->doSS && ind->solvedIdx >= idx condition that already distinguishes a
  fresh solve from a re-query (e.g. the output pass).
• linCmtBdoseTime() reads that cache back on a re-query instead of the
  by-then-cleared live rate.
• A steady-state infusion still returns NA_REAL: it establishes its amounts
  analytically (handleSSinf8()/solveSSinf()), and for a one-shot SS dose
  with no following schedule the solver deliberately leaves InfusionRate at
  0 afterward -- the cache-based value there is well-defined but not
  physically meaningful (verified empirically: true steady state should give
  -dA/dt = 0, but reading the cache at that index gave a nonzero answer).
  Steady-state bolus never touches InfusionRate and was already exact.

Both correctness bugs an independent Antigravity/Gemini review raised (a
stale ind->idx sentinel during extra-dose injection; dense dop853
intra-segment observations) were checked against the actual code and do not
apply here: the extra-dose/pushDosingEvent path is reachable only from
inside the steady-state branch of handle_evid, which the new
linCmtHasSsInfusion() guard already refuses regardless of cache state; and
dense output is unconditionally disabled for any model with numLin > 0
(rxData.cpp), i.e. any model that could call linCmtB() at all. Both were
confirmed unreachable by tracing the call graph and by passing tests that
exercise the surrounding code paths (steady-state infusions, method = "dop853").

Test plan

• tests/testthat/test-lincmt-dose-time-sens.R: infusion regimens (single,
  multiple, duration-specified, oral-into-depot, IV-with-peripheral, mixed
  linCmt() + d/dt(), method = "dop853") match a central finite
  difference to ~1e-10; steady-state infusions (periodic and continuous)
  still report NA.
• tests/testthat/test-lincmt-dose-time-sens-guard.R (unaffected linCmtB(which1 = -3) assumes every dose shares one alag(): no per-compartment dose-time sensitivity #1237
  guard) still passes.
• tests/testthat/test-lincmt-solve-sens.R, test-steady-state.R,
  test-event-sensitivities*.R, test-infusion-bolus.R,
  test-lincmt-*.R all pass with no new failures.
• Full clean rebuild (rm -f src/*.o, required since a shared header
  changed) with no new warnings from the changed files.
• Reviewed with an independent Antigravity (Gemini) pass; findings
  triaged above.