chore: modernize tests/CI and fix cross-platform/runtime bugs

[DiogoRibeiro7]

Summary

This PR improves repository reliability and developer workflow without changing model behavior.

What changed

• Modernized test workflow to pytest:
  • Added marker-based test taxonomy (integration, tensorflow)
  • Updated default test entrypoint (test_all) to run fast local tests
  • Added pytest.ini and test collection guards in conftest.py
• Added CI improvements:
  • New GitHub Actions workflow for unit tests across Python versions
  • Non-blocking TensorFlow smoke job
  • Updated Travis config to use the same test entrypoint
• Added offline, deterministic unit coverage:
  • New fixture-backed tests for metadata and feature utilities
  • New unit tests for GCSFile, metrics helpers, and config behavior
  • Added dependency-sentinel tests so marker-only runs skip cleanly
• Fixed cross-platform/runtime bugs:
  • GCSFile now closes file handles before temp cleanup (Windows-safe)
  • Removed Unix-only/unused imports that broke test collection on Windows
  • Fixed date parsing to timezone-aware UTC and removed deprecated datetime usage
  • Replaced Python 2 StandardError usage with Python 3-safe exception handling
  • Fixed broken helper functions in feature_utilities.py (undefined vars/offset logic)
  • Fixed runtime defects in metrics modules:
    • ydump.py missing imports and helper resolution
    • compute_vessel_metrics.py undefined labels bug in weights_by_class
    • compute_fishing_metrics.py unsafe debug prints and cleanup
• Packaging and repo hygiene:
  • Improved setup.py metadata/package discovery and dev extras
  • Added .gitignore entries for Python cache artifacts
  • Updated README with current test/marker execution guidance

Why

The repo had environment-coupled tests, legacy CI assumptions, and several latent runtime issues (including Windows-specific failures).
These changes make local development and CI predictable, improve portability, and prevent regressions through targeted tests.

Validation

• sh ./test_all passes
• python -m pytest -q passes/skips as expected in non-integration environments
• Added targeted unit tests for newly fixed behavior
• Ran lint/static checks on touched modules (targeted ruff/pyflakes)

Notes

• TensorFlow and GCS-backed integration tests remain opt-in by design.
• No intentional changes to model training/inference semantics.