[Metrics] Implement Cosine Similarity in Validation Pipeline

[DsThakurRawat]

Description

Currently, reflex validate and the core validate_outputs function only track absolute and relative differences (max_abs_diff, max_rel_diff).

Rationale

In 6DoF robotics, the direction of the action vector is often more critical than the absolute magnitude (especially for rotation). Cosine similarity is the industry-standard metric for verifying that a model's 'intent' remains aligned after ONNX/TensorRT optimization.

Tasks

• Add cos_sim to the ValidationResult dataclass in src/reflex/validate.py.
• Implement the cosine similarity calculation using np.dot / (norm_a * norm_b).
• Surface this metric in the VERIFICATION.md report.