Add evaluation script for trajectory prediction and completion tasks (Yasoz UniTraj)

[24daydayup]

Summary

Adapt an evaluation script for Yasoz/UniTraj based on the evaluate_downstream.py from 24daydayup/UniTraj, in order to properly evaluate the performance on trajectory prediction and completion tasks.

Background

• 24daydayup/UniTraj uses evaluate_downstream.py to evaluate MAE/RMSE for two tasks: trajectory prediction (预测) and completion (补全)，using batch-wise haversine distance calculation and compatible with their data/preprocessing format.
• Yasoz/UniTraj has its own transformer-based implementation (in utils/unitraj.py) and data pipeline (in utils/dataset.py).
• Currently, Yasoz repo lacks a corresponding downstream evaluation script that can (1) mask points in the input, (2) denormalize/restore lat/lon, and (3) calculate metrics compatible with convention (meters, prediction/completion mode etc).

What needs to be done

1. Create evaluate_downstream.py (or equivalent) in Yasoz/UniTraj, referencing the logic and metrics from 24daydayup/UniTraj:
   • Support both prediction (mask last N) and completion (mask arbitrary indices) evaluation modes.
   • Compatible with Yasoz/UniTraj's UniTraj model, data format, and normalization stats.
   • Batchwise metric (MAE, RMSE, meters, using haversine), support for torch DataLoader.
   • Output/save results and logs using Yasoz's logger or standard output.
2. Sample usage:
   • Load dataset via Yasoz-style TrajectoryDataset
   • Restore model and normalizer config
   • Evaluate model for both prediction/completion

Reference

• 24daydayup/UniTraj evaluate_downstream.py
• Yasoz/UniTraj: main.py, utils/unitraj.py, utils/dataset.py

Acceptance criteria

• Script can evaluate both prediction and completion tasks with proper masking scheme
• Metric output matches 24daydayup/UniTraj when using same model/dataset
• Results are logged/saved
• Short run/usage instruction included as docstring or comments

---

If desired, label as documentation (since it's mostly script adaptation + usage guidance)