Refactor the JACCL ring and add threads for the multiple rings

[angeloskath]

This is on top of the coordinator changes and I will rebase it when that is merged, for now focus on the changes in the reduction ops and more importantly the ring implementation.

The TL;DR is the following:

• Refactor the ring to extract the ring pass in a standalone function (with copy op or reduction)
• Refactor the reduction ops to support out of place reduction that allows us to avoid the initial copy
• Write a per-wire all reduce and all gather using the ring pass
• Copy the MLX threadpool and use it to dispatch the other wires to different threads

The performance on a single wire is maintained and there should be 0 overhead since that wire runs in the calling thread.

TODO: Add multi-wire benchmarks to show improved performance compared to before.