Benchmark new allocators in README figures and unify the 3 s budget

[fpedd]

Add best-fit, simulated annealing, tabu search, and telamalloc to the
README benchmark figures: the hero Pareto plot grows from 8 to 11
allocators (greedy-by-area retired; the size and all variants already
tell the greedy story), the quality plot now compares greedy (size),
best-fit, tabu search, telamalloc, minimalloc, and supermalloc, and the
scaling plot gains telamalloc (best-fit is omitted there because its
curve coincides with greedy). Two palette roles are added: amber for
the Python searchers and teal for telamalloc.

Unify every time-bounded allocator on one default wall-clock budget:
DEFAULT_MAX_SECONDS = 3.0 in allocators/base.py, mirrored by
kDefaultMaxSeconds in src/cpp/allocators/defaults.hpp. Simulated
annealing, tabu search, and telamalloc move from 2 s to 3 s;
supermalloc and minimalloc from 10 s to 3 s. HillClimbAllocator and
GeneticAllocator gain a max_seconds parameter (default 3 s, 0 disables
it): hill climbing checks a monotonic deadline each iteration, and the
genetic loop unrolls DEAP's eaSimple (verified RNG-identical) so it can
stop between generations. The figure script consequently drops its
per-allocator timeout plumbing and runs everything at library defaults.

All assets are regenerated from a fresh run: supermalloc proves
optimality on 10 of 13 hard problems within 3 s (down from 11 at 10 s)
at 97.6% mean efficiency, and mm-G in the README allocation figure
remains proven optimal. The 05 example now also exercises best-fit and
telamalloc.