Add pluggable trajectory aggregation with learned GBDT reference implementation

[lambdabaa]

Summary

• Introduces AbstractTrajectoryAggregator in base.py so aggregation is pluggable at the algorithm level, replacing hardcoded prod/min/mean reductions in ParticleFiltering and BeamSearch
• Ships HardcodedAggregator (wraps the three existing reductions), LearnedGBDTAggregator (recommended, sklearn), and LearnedMLPAggregator (torch, kept as an extensibility example)
• Adds MLXProcessRewardModel (Apple Silicon) and TransformersProcessRewardModel (includes RoPE buffer repair for transformers ≥5.0) to its_hub/integration/
• 231 new tests covering aggregators, algorithm integration, and both PRM implementations

Motivation

ParticleFiltering and BeamSearch previously hardcoded prod/min/mean with no override path. None of the three is a defensible default (prod penalises long trajectories, min is brittle to a single weak step, mean discards position). This PR makes the choice explicit and extensible.

Learned aggregator evaluation

Nested 5-fold CV across 408 mixed-difficulty MATH problems (Levels 1–5, N=8 trajectories per problem):

Aggregator	Mean acc	95% CI
random	0.380	[0.306, 0.454]
min	0.490	[0.419, 0.562]
prod / mean	0.498	[0.434, 0.577]
learned_mlp	0.495	[0.428, 0.562]
gbdt	0.517	[0.454, 0.581]

GBDT is the only model that consistently clears the fixed baselines across all 5 folds. Its decision surface concentrates ~78% of feature importance on min — effectively learning a continuous threshold on the worst step, which is the dominant signal at this data scale. MLP is statistically indistinguishable from prod/mean at 408 problems.

Interface

from its_hub.aggregators import HardcodedAggregator, LearnedGBDTAggregator

# Drop-in replacement — existing code unchanged (defaults to HardcodedAggregator("prod"))
pf = ParticleFiltering(sg=sg, prm=prm, aggregator=HardcodedAggregator("min"))

# Learned aggregator
agg = LearnedGBDTAggregator("its_hub/aggregators/checkpoints/gbdt_agg.pkl")
pf = ParticleFiltering(sg=sg, prm=prm, aggregator=agg)

# Custom aggregator
class MyAggregator(AbstractTrajectoryAggregator):
    def aggregate(self, step_scores: list[float]) -> float: ...

Test plan

• tests/test_aggregators.py — unit tests for all three aggregator classes, async delegate, and ParticleFiltering integration
• tests/test_algorithms.py — BeamSearch aggregator wiring
• tests/test_mlx_prm.py — MLXProcessRewardModel unit tests

[beatsmonster]

@lambdabaa could we please rebase to main/v1