Fix OOM in contrastive pair generation with streaming approach

[Wert1996]

Summary

This PR fixes out-of-memory issues when training SetFit on large datasets by replacing eager O(n²) pair generation with streaming pair generation.

Problem

Training with contrastive loss (e.g., CosineSimilarityLoss) on datasets with a large enough number of samples causes OOM before training even starts. The root cause is three layers of O(n²) memory allocation:

1. shuffle_combinations() creates all pair indices upfront
2. ContrastiveDataset stores all pairs in pos_pairs/neg_pairs lists
3. Dataset.from_list(list(...)) materializes the iterator again

Solution

sampler.py:

• Replace shuffle_combinations() with on-the-fly random pair sampling
• ContrastiveDataset now stores only label_to_indices mapping (O(n))
• __iter__() generates pairs on-the-fly with set-based uniqueness tracking
• Same changes for ContrastiveDistillationDataset

trainer.py / trainer_distillation.py:

• Use IterableDataset.from_generator() instead of Dataset.from_list(list(...))
• Compute max_steps automatically for IterableDataset compatibility

Memory Comparison

Component	Before	After
Index arrays	O(n²)	0
Pair lists	O(n²)	0
Label grouping	0	O(n)
Uniqueness set	0	O(num_pairs)
Dataset copy	O(n²)	0

Breaking Changes

• ContrastiveDataset.pos_pairs and neg_pairs attributes removed
• ContrastiveDataset.len_pos_pairs / len_neg_pairs now represent target counts, not stored counts
• Added estimated_num_pairs property for logging

Testing

• Verified training completes on large datasets that previously OOM'd
• Pairs maintain uniqueness via set-based tracking
• Reproducibility preserved via seeded RNG