implement contrastive loss

[tashapais]

I've successfully implemented contrastive loss for PufferLib with the following key components:

Core Implementation (pufferlib/contrastive_loss.py):

• ContrastiveLoss class with InfoNCE loss using geometric future sampling
• Functional interface compute_contrastive_loss_pufferlib() for easy integration
• Proper handling of episode boundaries using terminal/truncation flags
• Metrics tracking for logging (positive/negative similarities, number of pairs, etc.)

Integration Example (pufferlib/pufferl_with_contrastive.py):

• Shows how to extend PufferLib's training loop to include contrastive loss
• Demonstrates proper embedding extraction and loss computation
• Maintains compatibility with existing PPO losses

Key Features:

• ✅ Samples random (st, at) pairs from replay buffer
• ✅ Creates positive examples sf^(1) using geometric distribution Δ ~ GEOM(1-γ)
• ✅ Generates negatives by shuffling future states from other trajectories
• ✅ Uses unnormalized representations as specified
• ✅ Integrates seamlessly with PufferLib's architecture
• ✅ Tested and working with synthetic data

The implementation is ready to use - you can integrate it into your training by importing the contrastive loss
function and adding it to your training loop as shown in the integration example.