Normalize pyro-dataset imports via dvc import

[Chouffe]

Summary

• Standardize how experiments import the pyro-dataset: always use dvc import from https://github.com/pyronear/pyro-dataset with a pinned version tag (v2.2.0)
• Convert tracking-fsm-baseline and mtb-change-detection from S3-only .dvc files to proper dvc import with a truncate pipeline stage for the 20-frame-max preprocessing
• Add "Dataset Import" section to experiments/GUIDELINES.md and update the experiment template with import instructions and a truncate stage scaffold

tracking-fsm-baseline changes

• Add truncate stage to dvc.yaml (import full dataset to datasets_full/, truncate to datasets/)
• Update list_sequences to handle nested {wildfire,fp}/ layout (matching pyro-detector-baseline)
• Update is_wf_sequence to use parent directory name instead of label format heuristic
• Add find_sequence_dir helper, used in track.py, sweep.py, ablation.py, build_fiftyone_dataset.py
• Fix load_label_boxes float class_id parsing (int(float(...)))
• Update tests for new is_wf_sequence, add tests for find_sequence_dir and list_sequences

mtb-change-detection changes

• Add truncate stage to dvc.yaml
• Replace prepare_dataset.py with truncate_sequences.py (single-split interface for DVC foreach)

Metrics comparison

Metrics are consistent with baseline (slight differences from complete v2.2.0 dataset having a few more sequences).

tracking-fsm-baseline val/pyronear: P=0.926 R=0.966 F1=0.945 FPR=0.060
mtb-change-detection val/pyronear: P=0.682 R=0.922 F1=0.784 FPR=0.331 (exact match)

Test plan

• make lint + make format pass for both experiments
• make test passes (106/106 tracking-fsm, 88/88 mtb)
• uv run dvc repro evaluate produces correct metrics for both experiments
• make fiftyone builds datasets successfully (tracking-fsm-baseline)
• New .dvc import files have frozen: true and repo.rev: v2.2.0