[Filestore] TStorageServiceActor overhead optimization

[qkrorlqr]

There are at least 2 big issues:

• there's a bug because of which TStorageServiceActor is spammed with EvPingSessionRequests - we see that for high iops tests (high iodepth and/or numjobs and small request size) TStorageServiceActor is 100% busy and 20% of that time is spent processing EvPingSessionRequests
• another 30% is spent in TStorageServiceActor::CompleteRequest which mostly does statistics accounting - this thing can be moved to a separate actor which wouldn't be on the critical path for IO processing

If we don't hit any other bottlenecks, freeing half of the time that TStorageServiceActor currently spends processing pings and stats to let it do IO instead may let us increase our max iops by x2.

A simple fio config which I used to trigger the issue:

[global]
filesize=2G
time_based=1
startdelay=5
exitall_on_error=1
create_serialize=0
filename_format=$filenum/$jobnum
group_reporting=1
clocksource=gettimeofday
ioengine=libaio
disk_util=0
direct=1

[init]
blocksize=1Mi
rw=write
size=100%
numjobs=64
time_based=0
description='pre-create files'

[read-4k]
stonewall
description='Read iops workload'
iodepth=8
bs=4k
rw=randread
numjobs=64
runtime=120

Another interesting thing which is worth testing here is how we scale if we increase guest queue count and max_background.