- Add environment variables to prevent over-allocation of GPU VRAM
  by TensorFlow and/or Jax.
- Add 'py{3x}-ci' test environments that are tweaked specifically
  for our CI runner, adding instructions that offer as much support
  for GPU as possible given our runner host's configuration.
- At the moment, an update of the host to the latest CUDA version
  enables native support for Torch, and TensorFlow with Python>=3.9
  (but not 3.8), plus Jax with some extra installation command.