Fix/correctness - #5

1. popen.maxPopen - five errors in the non-monotonic bisection loop

   A plain agonist-gated mechanism gives a monotonically rising Popen curve,
   so this branch is never entered and the bugs stayed hidden. Adding a fast
   pore blocker makes the curve rise and then fall, which exposes them:

     a. P1 was evaluated at conc instead of conc/fac
     b. conc1 was overwritten by conc*fac, destroying the lower probe point
     c. P2 was evaluated at conc instead of conc*fac
     d. the bracket updates were backwards and used conc1, not the midpoint
     e. nstep was never incremented

   Net effect: P1 == P2 always, so perr == 0 and the loop exited after one
   iteration with no real bisection. Measured on CH82 with a 1 mM fast
   blocker, maxPopen returned 0.90337 where a dense scan finds 0.92695 - a
   2.6 % error in the reported maximum, at the wrong concentration.

   The loop now probes either side of the midpoint, moves the correct bracket
   end, terminates on the concentration tolerance, and returns the midpoint of
   the final bracket.

2. AsymptoticPDFCalculator._calculate_roots - overflow at large tres

   The lower search bound was fixed at -1e6, so |s| * tres exceeded
   ln(float64 max) ~ 709 once tres was above about 0.7 ms.
   exp((QFF - s*I)*tres) then overflowed and inf/nan reached W(s):

     numpy.linalg.LinAlgError: Array must not contain infs or NaNs

   Reproduced with CH82: fine at tres = 700 us (|s|*tres = 700), raises at
   1 ms (1000). The bound is now clamped to -700/tres, and only when the
   default would overflow, so ordinary resolutions keep the full search range.

3. AsymptoticPDFCalculator._bisect_intervals - infinite loop

   Subintervals containing no roots were pushed back onto the work list, so
   they split into two more empty halves for ever. Mechanisms with a single
   open state never returned: CO and CCO both hang, at tres = 0 and at
   tres = 100 us. Zero-root subintervals are now discarded.

Tests
-----
tests/test_popen_maxpopen.py (8 tests) and tests/test_hjc_roots.py (14 tests).
Against the unfixed code 2/8 and 9/14 fail respectively; all pass after the
fix. The five hjclib tests that pass either way are the CH82 multi-open-state
cases, which neither bug affects.

The root-finding tests run each call in a subprocess with a timeout, so a
regression of bug 3 reports a test failure instead of hanging the run.

Co-Authored-By: Claude Opus 5 <[email protected]>