Fix vLLM compatibility in GreedyProbsCalculator and WhiteboxModelvLLM

[dmndxld]

Summary

Fix two bugs preventing vLLM models from working with estimate_uncertainty() and UEManager.

Changes

1. GreedyProbsCalculator vLLM compatibility (stat_calculators/greedy_probs.py)

• Skip output_attentions and output_hidden_states when model is vLLM (not supported by vLLM engine)
• Skip embeddings extraction and attention processing for vLLM models
• Without this fix, using vLLM with any estimator triggers NotImplementedError in get_embeddings_from_output()

2. WhiteboxModelvLLM missing model_path attribute (model_adapters/whitebox_model_vllm.py)

• Call super().__init__(model_path, model_type) to properly initialize base class
• Add model_path parameter (defaults to vLLM engine's model_name)
• Without this fix, estimate_uncertainty() raises AttributeError: 'WhiteboxModelvLLM' object has no attribute 'model_path'

Test plan

• Verified with estimate_uncertainty() using WhiteboxModelvLLM + SemanticEntropy
• flake8 and black checks pass locally