Parent wait loop: remove the 10 ms usleep polling floor on fast runs #224

Hub: #15. Related: #217 / #220.

Problem

The parent process polls the solver child in solve_in_child_fork (src/core/distqldpc.cc):

for (;;) { drain pipe; waitpid(pid, WNOHANG); deadline check; usleep(10000); }

The parent only notices that the child has exited at the next 10 ms tick. Fast runs therefore get a reported wall time quantized into 10 ms buckets: LP_442 with -engine=cc lands at about 23 ms or about 33 ms. In PR #220's quiet interleaved Tier 1 panel, candidate A uses about 9 % more CPU on LP_442, which moves every A run into the upper bucket. The result is a spurious 1.467x cell. This floor affects every engine (MaxSAT and CC), not only CC.

Plan

• Smallest sound change: the parent sleeps in poll() on the bounds pipe and on a SIGCHLD self-pipe. The timeout is the remaining -cpu-lim time, capped at 100 ms as a safety net. The child's exit wakes the parent immediately, and bound lines are relayed as soon as they arrive.
• Keep the time limit (-cpu-lim wall deadline -> SIGKILL -> TIMEOUT with bounds), the memory cap (prlimit on the process tree) and the kill behaviour unchanged. The solver child's code and output are untouched.
• Tier 0: make test-cc, make test-cc-flags, smoke test; answers and logs must be unchanged apart from timing.
• yfclab2: Tier 0 + Tier 1 (layout of 972b28c / fecee8c), 5 seeds x {MaxSAT default cardinality, MaxSAT -no-card, CC default}, old vs new binary interleaved on one core per cell. Pass = geomean < 1 and no cell > 1.25. Spot-check a handful of Tier 2/3 codes; Tiers 2-5 are unaffected by construction (the solver is unchanged; the parent only wakes less often).
• Then re-run GH-217: CC default rules + tiebreak, sparse rows and -cc-roworder=auto (Tier 1 floor: PI decision) #220's Tier 1 interleaved panel (A rebased on the new main vs the new main).