This PR adds a numba implementation of the loop in SugarscapePolars to get the best moves for each of the agent, avoiding the (slightly) more complex reasoning of the DF Loop.

Some observations so far:

• Looping DF is still the fastest, followed by numba and the completely iterative (non vectorized) process (see polars_comparison.png)
• Working with numba is a bit tedious as:
  • you have to specify types and sizes of vectors
  • must have an input vector which is the size of the output
  • must use 1d vectors
• Most of the simulation time is spent on join operations, outside of getting the best moves, mostly on getting the neighborhood and setting the cells (which could be sped up with a lazy approach and a faster backend overall). Results in the flame graph polars_comparison.svg

I am still trying to get the cuda target to work, will update.
Also the non vectorized loop could potentially be faster if you iterate directly on the polars DF but the point was to check how much worse non vectorization would have been.