Initial version of a custom ring reduce backend. It can be used with Ethernet as well but the main focus is on thunderbolt because:

1. MPI ring reduce is not a very good fit and doesn't saturate the TB interface.
2. it makes it much easier to add more cables for even more bandwidth

The comparison is as follows for a 4-way all reduce using 4 M2 Ultras:

The bandwidth is computed as total bytes * 3/4 * 2 / time the 3/4 is due to the ring reduce actually sending less than 2*B bytes on the wire.

In terms of time per all reduce

which means that for 2GB all reduce we go from 1.3 seconds over Ethernet and 0.45s with MPI to 0.27s with our reduce.

Things to note:

1. mx.distributed.init() takes a backend argument and initializes the corresponding backend. If no argument is provided it tries to initialize in order ring > mpi. The reason is that ring will fail if MLX_HOSTFILE and MLX_RANK is not provided while mpi may succeed and "hide" the rest of the backends.
2. mx.distributed.init() called for a 2nd time without an argument returns the previously initialized backend. This allows the rest of the code to remain unchanged and agnostic of the communication group.
3. Small edits to the Threadpool to make it resizable.
4. The ring backend can be improved/simplified so take this as a 1st version mostly defining the API.

Finally, I am in the process of preparing another PR with changes in the docs and possibly a launcher for this backend so people don't have to write their own shell scripts to launch with this backend.