TODO-coul

TODO (high priority):
- calculate even more accurate mintau()
  - and beyond oul for suppressed factors
- fix linear search in pell.c:contfrac()
- 'coul -p20 -x1e80 66 5' spends all its time finding allrootmod()
  results, almost all of which are immediately greater than endr
- when n=36 and we allocate p^5, the remainder cannot allocate p^2 so
  we can require a 5th power. But our "have_square" support is predicated
  on the assumption of a square: how readily can we extend it to cater
  for this case? This case kicks in when (x > t/2 && x.high == t.high),
  and will be useful for n in { 36, 54, 72, 100 }. n=18 is a more extreme
  case.

CHECKME:
- inject failures can clash with walk_1() successes
- can best_v do better by picking the most constrained?

TODO (lower priority):
- add parallelization support to walk_1
- maintain a dummy alloc[0] on values[], to save lots of special handling
  - this may remove the need for m[] and t[] in walk_v()
- short-circuit "used" check with cur->maxp
- on recover, parse 001 line to check consistency of critical options
- reimplement small_divmod()
- reimplement next_prime()
- cater for end value > 2^64 in walk_v, if needed
- add support for -o
  - build it over parallel is_tau() checks, so it can be configurable
- prep the simple_valuation() checks for apply_batch()
- for unforced primes, do constant apply_alloc() work in advance
- avoid recursing when best_v() will find nothing