Improve speed and parallelization for raisimpy gym

[bdelhaisse]

Status

The current code for raisimpy_gym is slower than the wrapped C++ code for raisimGym.
There are several ways to improve the performances.

The problem with threads

Currently, the Python code uses threads however due to the Global Interpreter Lock (GIL), the threads are prevented to run truly in parallel. Worse than that, there are some overheads due to acquiring and releasing the lock, thus it is actually better to not use threads.

There are 2 solutions (that I can currently think of) that could improve the performances:

• using pybind11: use threads in C++ that call an arbitrary Python function. As the documentation is scarce (GIL in pybind11), I will need to test this, and this might actually not work.
• using multiple processes instead of threads: each process has its own GIL so by using multiple processes you don't have the problems with threads. There are several multiprocessing libraries that could be useful: multiprocessing, pathos.multiprocessing, joblib, concurrent.futures, and ray. For the moment, I quickly tested and could make the current code works with concurrent.futures.ProcessPoolExecutor, which improved a bit the performances compared to the single thread/process code.

Copy vs Reference

Currently, I am passing by value the numpy arrays and Eigen matrices/vectors instead of passing them by reference, and so there is a copy everytime we convert between these 2 datatypes, as described here: https://pybind11.readthedocs.io/en/stable/advanced/cast/eigen.html?highlight=eigen
The solution is to pass them by reference instead, and make sure for that for matrices we pay attention to the fact that Eigen uses a column-major order, while numpy uses a row-major order.