Implement context parallelism (RingAttention) as `KVCache`

**Is your feature request related to a problem? Please describe.**

LARGER PROJECT

Currently, all `KVCache` implementations store buffers and run computation on a single device. RingAttention distributes memory and compute across several devices. Our `KVCache` abstraction allows for that to be implemented fairly easily.

**Describe the solution you'd like**

To the user, this should look like a normal KV cache, just with a much larger length:

* Different to RingAttention: Any KV cache type supported on the devices (but same type on each)
* Allows for much larger cache length
* RingAttention also parallelizes over the query axis. So this speeds up processing of large chunks

Should be doable with `isend`, `irecv`: https://docs.pytorch.org/tutorials/intermediate/dist_tuto.html#our-own-ring-allreduce

Approach:
* Copy KV buffers to CPU and transfer them from there. Sync point at the end of the computation
* Each rank computes part along query axis
* At the end: all_reduce to spread the results

Note: We cannot couple this with DDP, to obtain a larger batch size. RingAttention parallelizes over the query axis as well, but the batch size also scales the KV cache buffers. We can get away with larger chunks here, though.