fixes #560: add computational cost tab to GUI evaluator

[AdityaX18]

Summary

Fixes #560

Adds a Computational Cost tab to the Streamlit GUI that wraps TorchImageDetectionModel.get_computational_cost() for the currently loaded detection model. Closes the GUI ↔ CLI feature parity gap flagged in the discussion on #539.

Closes #XXX.

Changes

• New tabs/computational_cost.py — standalone tab function following the *_tab() convention used by the other three tabs. No shared base (matches the existing codebase pattern).
• Modified app.py — three edits:
  1. Import computational_cost_tab
  2. Extend st.tabs([...]) from 3 to 4 entries
  3. Extend the PAGES dict to stay in sync

How it works

• Reuses the same sidebar-loaded model via st.session_state.detection_model (consistent with evaluator_tab and inference_tab).
• Gates the expensive get_computational_cost() call behind st.button("Run Cost Analysis") so widget interactions (number inputs) don't retrigger 30+ forward passes on every keystroke.
• Persists the resulting DataFrame in st.session_state.computational_cost_result so re-renders on widget changes reuse the cached result instead of recomputing.
• Renders metrics using the same unit conventions as examples/torch_computational_cost.py (n_params / 1e6 for M, inference_time_s * 1000 for ms) — no new formatting invented.
• Derived throughput (FPS = 1 / inference_time_s) added as an extra metric — not in the DataFrame, but useful for deployment decisions. Handles the zero-division edge case.
• CSV export via st.download_button mirrors the CLI's results.to_csv(out_fname) behavior.

Testing

Manual, locally:

• streamlit run app.py launches without errors.
• All three pre-existing tabs still render correctly.
• New tab shows a warning when no detection model is loaded.
• After loading a torchvision detection model from the sidebar, the tab accepts custom image size + runs + warm-up and produces sensible numbers (params in M, size in MB, latency in ms, FPS derived correctly).
• CSV download produces a file with the same 4-column schema the CLI produces.
• Modifying the number inputs after a run does NOT retrigger the forward passes (button-gated, confirmed via terminal output).

Out of scope

• No changes to get_computational_cost() itself. FLOPs / MACs reporting is proposed as a follow-up issue with a separate dependency discussion.
• No segmentation / LiDAR changes. The GUI currently supports detection only per the README; extending to other tasks is a follow-up when GUI support is extended.