Add NeuralAllocator, a learned score-and-sort allocator

[fpedd]

A permutation-equivariant Transformer (~360k params, PyTorch, optional
"ml" extra with a CPU-only wheel source) scores every allocation of a
problem; placing the allocations first-fit in descending score order
yields the full solution, since the placement order alone determines
the outcome. The scores define a Plackett-Luce distribution over
permutations, so inference takes the best of the deterministic decode,
the classic greedy orders (portfolio floor: never worse than
greedy_by_all), and up to 1024 Gumbel-perturbed policy samples within
the shared 3 s budget.

Features are scale-invariant per instance: normalized sizes, times,
areas, conflict degrees, their ranks, plus instance-level context
(each classic greedy order's relative peak and a winner one-hot)
broadcast to every allocation.

Training (training/train_neural.py) runs on CPU in under an hour:
Plackett-Luce behavior cloning of the tie-consistent best greedy order
per instance, then self-imitation (expert iteration) toward
per-instance incumbent orders found by policy sampling, with peak
memory evaluated by the C++ FirstFitPlacer and normalized by the
max-live-load lower bound. Problems are freshly generated from the
generator sources with randomized parameters; the minimalloc CSV
datasets stay held out for evaluation. The fp16 checkpoint (487 KB)
ships with the package.

On 90 held-out problems (unseen generator seeds, minimalloc small and
challenging) the mean peak over lower bound is 1.058, ahead of every
heuristic and metaheuristic (greedy_by_all 1.080, simulated annealing
1.082, telamalloc 1.083, hillclimb 1.146) and behind only the exact
supermalloc solver (1.008); training/eval_neural.py reproduces the
comparison, including decode-only and sampling-only ablations.