Mpsc channel v3

[mratsim]

A rewrite of the core MPSC channel with the following benefits:

• No CompareExchange at all for the consumer. CompareExchange is the costliest atomic especially on weak memory model
• No spinlock / cpuRelax. While they don't yield the processor, busy waiting is bad style.

This is the revival of the v1 channels that didn't support batching, but with a State Machine twist: https://github.com/mratsim/weave/blob/b796d1a0efb15cad68e2b06ddaa305b08cf5f722/weave/channels/channels_mpsc_unbounded.nim

Trying to add batching support to the tryRecv proc was incredibly hard and ultimately failed due to the moving dummy and the non-trivial control flow, hence the v2 design with a fixed dummy node in front.

Pure queue benchmarks are inconclusive, the v3 seems to be slightly better for single task dequeue but is slower for batch dequeue :/.

Impact on actual workload to be measured

[mratsim]

10% overhead measured on fib(40), no impact on matmul.

Unless somehow the deadlock/livelock that happens regularly in CI don't happen, stashing this failed experiment