kvserver: shard Raft scheduler

The Raft scheduler mutex can become very contended on machines with many
cores and high range counts. This patch shards the scheduler by
allocating ranges and workers to individual scheduler shards.

By default, we create a new shard for every 16 workers, and distribute
workers evenly. We spin up 8 workers per CPU core, capped at 96, so 16
is equivalent to 2 CPUs per shard, or a maximum of 6 shards. This
significantly relieves contention at high core counts, while also
avoiding starvation by excessive sharding. The shard size can be
adjusted via `COCKROACH_SCHEDULER_SHARD_SIZE`.

This results in a substantial performance improvement on high-CPU nodes:

```
name                                    old ops/sec  new ops/sec  delta
kv0/enc=false/nodes=3/cpu=4              7.71k ± 5%   7.93k ± 4%     ~     (p=0.310 n=5+5)
kv0/enc=false/nodes=3/cpu=8              15.6k ± 3%   14.8k ± 7%     ~     (p=0.095 n=5+5)
kv0/enc=false/nodes=3/cpu=32             43.4k ± 2%   45.0k ± 3%   +3.73%  (p=0.032 n=5+5)
kv0/enc=false/nodes=3/cpu=96/cloud=aws   40.5k ± 2%   61.7k ± 1%  +52.53%  (p=0.008 n=5+5)
kv0/enc=false/nodes=3/cpu=96/cloud=gce   35.6k ± 4%   44.5k ± 0%  +24.99%  (p=0.008 n=5+5)

name                                    old p50      new p50      delta
kv0/enc=false/nodes=3/cpu=4               21.2 ± 6%    20.6 ± 3%     ~     (p=0.397 n=5+5)
kv0/enc=false/nodes=3/cpu=8               10.5 ± 0%    11.2 ± 8%     ~     (p=0.079 n=4+5)
kv0/enc=false/nodes=3/cpu=32              4.16 ± 1%    4.00 ± 5%     ~     (p=0.143 n=5+5)
kv0/enc=false/nodes=3/cpu=96/cloud=aws    3.00 ± 0%    2.00 ± 0%     ~     (p=0.079 n=4+5)
kv0/enc=false/nodes=3/cpu=96/cloud=gce    4.70 ± 0%    4.10 ± 0%  -12.77%  (p=0.000 n=5+4)

name                                    old p95      new p95      delta
kv0/enc=false/nodes=3/cpu=4               61.6 ± 5%    60.8 ± 3%     ~     (p=0.762 n=5+5)
kv0/enc=false/nodes=3/cpu=8               28.3 ± 4%    30.4 ± 0%   +7.34%  (p=0.016 n=5+4)
kv0/enc=false/nodes=3/cpu=32              7.98 ± 2%    7.60 ± 0%   -4.76%  (p=0.008 n=5+5)
kv0/enc=false/nodes=3/cpu=96/cloud=aws    12.3 ± 2%     6.8 ± 0%  -44.72%  (p=0.000 n=5+4)
kv0/enc=false/nodes=3/cpu=96/cloud=gce    10.2 ± 3%     8.9 ± 0%  -12.75%  (p=0.000 n=5+4)

name                                    old p99      new p99      delta
kv0/enc=false/nodes=3/cpu=4               89.8 ± 7%    88.9 ± 6%     ~     (p=0.921 n=5+5)
kv0/enc=false/nodes=3/cpu=8               46.1 ± 0%    48.6 ± 5%   +5.47%  (p=0.048 n=5+5)
kv0/enc=false/nodes=3/cpu=32              11.5 ± 0%    11.0 ± 0%   -4.35%  (p=0.008 n=5+5)
kv0/enc=false/nodes=3/cpu=96/cloud=aws    14.0 ± 3%    12.1 ± 0%  -13.32%  (p=0.008 n=5+5)
kv0/enc=false/nodes=3/cpu=96/cloud=gce    14.3 ± 3%    13.1 ± 0%   -8.55%  (p=0.000 n=4+5)
```

Furthermore, on an idle 24-core 3-node cluster with 50.000 unquiesced
ranges, this reduced CPU usage from 12% to 10%.

The basic cost of enqueueing ranges in the scheduler (without workers or
contention) only increases slightly in absolute terms, thanks to
`raftSchedulerBatch` pre-sharding the enqueued ranges:

```
name                                                                 old time/op  new time/op  delta
SchedulerEnqueueRaftTicks/collect=true/ranges=10/workers=8-24         457ns ± 2%   564ns ± 2%   +23.36%  (p=0.001 n=7+7)
SchedulerEnqueueRaftTicks/collect=true/ranges=10/workers=16-24        461ns ± 3%   563ns ± 2%   +22.14%  (p=0.000 n=8+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=10/workers=32-24        459ns ± 2%   591ns ± 2%   +28.63%  (p=0.000 n=8+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=10/workers=64-24        455ns ± 0%   776ns ± 5%   +70.60%  (p=0.001 n=6+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=10/workers=128-24       456ns ± 2%  1058ns ± 1%  +132.13%  (p=0.000 n=8+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=100000/workers=8-24    7.15ms ± 1%  8.18ms ± 1%   +14.33%  (p=0.000 n=8+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=100000/workers=16-24   7.13ms ± 1%  8.18ms ± 1%   +14.77%  (p=0.000 n=8+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=100000/workers=32-24   7.12ms ± 2%  7.86ms ± 1%   +10.30%  (p=0.000 n=8+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=100000/workers=64-24   7.20ms ± 1%  7.11ms ± 1%    -1.27%  (p=0.001 n=8+8)
SchedulerEnqueueRaftTicks/collect=true/ranges=100000/workers=128-24  7.12ms ± 2%  7.16ms ± 3%      ~     (p=0.721 n=8+8)
```

Epic: none
Release note (performance improvement): The Raft scheduler is now
sharded to relieve contention during range Raft processing, which can
significantly improve performance at high CPU core counts.