Adds native vector implementation

[pseudo-rnd-thoughts]

Projects like SampleFactory and EnvPool have provided C++ based vectorisators for ALE, significantly increasing the number of frames per second that can be sampled.

However, EnvPool is no longer maintained and to make it easier and compatible with Gymnasium's vector API, this PR adds a EnvPool based (though highly highly simplified) vector environment.

The implementation has four main parts

1. preprocessed_env.hpp - Runs the standard preprocessing computation for Atari (frame stack, frame skip, maxpooling, resizing, etc)
2. async_vectorizers.hpp - Takes a list of preprocessed environments that are run asynchronously that we push actions to then return a collection of results once computed.
3. ale_vector_python_interface.cpp/hpp - PyBind11 bindings to convert the vectorizer's list of environment data into a batched version that is easier to use by users
4. vector_env.py - This is a python class for easily using the vectorizer.

Known Problems

• For MacOS ARM (not x86), the ALE vector and AtariPreprocess are NOT equivalent for all timesteps, however, the difference in only 1 or 2 pixels with a difference of 1. The issue is almost certainly with OpenCV with a different implement between MacOS ARM and x86.

Features not included in the PR (todo later)

• This PR only supports next-step autoreset mode, future PRs could look to add disabled autoreset mode (relatively easy) and same-step autoreset mode (more difficult)
• XLA support - The EnvPool implementation adds support for jit accelerating the environment, however, the jax foreign function interface has changed significantly since the EnvPool therefore will need to be implemented from scratch.
• Testing - This PR is minimally tested, particularly checking that agents can be successfully trained

[jjshoots]

Reviewing it soon, thanks for the work on this!

[jjshoots]

Admittedly, a lot of it flew over my head and I never ran the code. But thus doesn't seem to break any existing functionality so I thibk it's good to go. Any downstream bugs can be fixed as they come a la open source.

[pseudo-rnd-thoughts]

@jjshoots Thanks for the review, it was complex code and took me a long time to implement so don't worry if you don't fully understand