Schedule SuperMalloc rounds under node budgets and speed up its search

[fpedd]

With fewer than six threads the portfolio ran only the incumbent bound and its members one after another, so the first member consumed the whole budget and machines with one to four cores solved none of the eleven challenging Minimalloc instances within 3 s. Rounds now always include the lower-bound rung and run under doubling node budgets, try_solve_many reports whether a member stopped early so an empty result stays a proof, and nine of the eleven are now proved optimal on one thread in under a second each, with no regression at 8 or 16 threads. The search itself re-keys its candidate trees only for placements that pass the bound check and marks placed allocations with a sentinel, raising node throughput two to nine times with byte-identical search trees, decomposition buckets allocations once instead of scanning all of them per band (128k allocations took 15.6 s, now 0.26 s), members are copied once inside the worker instead of three times, the reorders are built only when greedy misses the lower bound, and the deadline is checked from the first node. Random 100k instances drop from 0.9 s to 0.35 s and instance D reaches a lower peak within the default budget; a brute-force and reference fuzz over about 35000 instances and every flag combination found no mismatch. The SuperMalloc integration test checks proved optima on challenging instances instead of running a 40 s benchmark campaign. Stacked on #75.