Add O(N log N) sweep and hybrid sweep allocators

[fpedd]

Chronological sweep placement for very large problems: allocation/free
events are processed in time order over an address-ordered, coalescing
free list (C++ sweep_place), with first-fit, best-fit, and two-ended
(median split) fit policies. Hybrid variants place the largest
max_obstacles allocations with the exact quadratic first-fit and sweep
the rest around them as forbidden offset bands (hybrid_sweep_place);
with the budget fixed at its default of 1024 the family stays
O(N log N). SweepByAllAllocator runs all 7 variants in parallel and
keeps the best result, mirroring GreedyByAll.

On a 33-problem suite (minimalloc datasets, synthetic sources, ONNX
models) the portfolio matches the quadratic GreedyByAll baseline at
geomean 0.9967 (2 wins, 29 ties, 2 losses, worst +1.7%) while placing
1M allocations in 0.83 s for a single sweep versus hours extrapolated
for one quadratic greedy variant. Hybrid runs with max_obstacles >= N
are bit-identical to the corresponding exact greedy variants, used as
a test invariant.

Total sizes beyond the free list's 2^61 top-gap bound raise
OverflowError instead of overflowing offset arithmetic, and a negative
max_obstacles raises ValueError. Research notes, prototypes, and the
quality benchmark used to select the shipped variants live in
research/, excluded from ruff/ty as a scratch area.