Harden regression workflow dependency install

[eaton-lab]

Summary

This hardens the regression test workflow after the recent PR exposed some fragility in the CI environment
setup.

The current workflow can fail during dependency installation because it installs system packages without
refreshing apt metadata, relies on some transitive Python dependencies for the documentation scenarios, and
always uploads the same artifact name across the test matrix.

Changes

• enable pip caching in actions/setup-python
• run apt-get update before installing system packages
• install ffmpeg and ghostscript with -y --no-install-recommends
• use python -m pip consistently
• install nbformat explicitly, since the documentation regression step imports it directly
• keep the existing .[all] install path for now
• upload failed-test artifacts only when a job fails
• give failed-test artifacts a matrix-specific name to avoid collisions

Why

This is a workflow-only fix. It does not change Toyplot runtime behavior.

The goal is to make the regression job more reliable across the Python version matrix and reduce avoidable CI
failures caused by environment setup rather than code regressions.

Validation

This PR changes only .github/workflows/regression-tests.yml.

I verified the workflow diff locally with:

git diff --check -- .github/workflows/regression-tests.yml