Refine MAPPO training and actor export Learning Path

[armwaheed]

Context

This is a follow-up to #3680 and pairs with the Device Connect dashboard continuation in #3694.

@spsagar13: these are intentionally broad review changes rather than a judgment on your contribution. Please feel free to reject or close this PR if the scope creates too much work for you.

Summary

• narrow the Learning Path to one train, evaluate, and export workflow
• remove the unavailable companion GUI page; Device Connect is the GUI in Part 2
• pin the BenchMARL revision and PyTorch, TorchRL, TensorDict, and VMAS versions
• prevent a zero-worker configuration and reduce nested thread-pool oversubscription
• add quantitative first, final, and best evaluation-return reporting
• replace whitespace-sensitive checkpoint discovery and shell-to-Python path interpolation
• reduce the export page from 594 lines by serving the reviewed script as a page resource
• make the exporter fail closed on absent deployment metadata and frame counts
• add source identity to output names, refuse accidental overwrites, validate before atomic replacement, and omit absolute host paths from metadata
• add recaps and a direct handoff to the Device Connect Learning Path

Validation

• hugo --minify
• hugo --buildDrafts --minify
• confirmed the exporter is emitted as a Hugo page resource and the removed GUI page is absent
• python3 tools/verify_index_fields.py ...
• all Markdown Python blocks compile and Bash blocks pass bash -n
• python3 -m py_compile export_mappo_actor.py
• ruff check export_mappo_actor.py
• ruff format --check export_mappo_actor.py
• focused tests for required configuration, atomic output, successful validation, and cleanup after validation failure
• confirmed Python 3.12 arm64 wheels for the pinned PyTorch, torchvision, torchaudio, TorchRL, and TensorDict versions; confirmed the pinned VMAS source distribution

Remaining validation

I did not rerun the full 330-minute training workload. The Learning Path and its paired Part 2 remain drafts; the training and export workflow should be rerun on the target Arm Ubuntu 24.04 instance before either draft is published.