Alternative coalescing for fast disks

[robert3005]

Random access benchmark improves by 30% if you set coalesce threshold to 0, i.e. never coalesce unless the buffers are adjacent which suggests that on nvme disks we would prefer to issue more smaller read requests instead of coalescing.

Additionally we assume that coalescing is beneficial and collect all the read requests into one shared queue that are drained by the io driver. Those channels can be slow if all you really want to do is get a byte range from a file. If we aren't going to coalesce we could come up with a simpler model where every cpu task could have its own io queue

[gatesn]

We have also always held that we need to gather latency + throughput statistics to tune the coalescing window