[hubcio]

The consensus tick visited every partition of a shard every 10 ms,
and a visit of an idle partition found nothing to do. Idle CPU
grew with the partition count: about 350% of one core per node at
30,000 partitions.

A partition now sleeps while it is quiet: Normal status, nothing
to commit, replicate, persist, repair or transfer, and no report
or gauge that waits for a visit. The partition table keeps one
wake tick per slot, the earliest timer deadline of the partition.
Every accessor that hands out a partition wakes it, so a frame, an
I/O turn or a reconcile step needs no call of its own. A partition
that is not plainly idle stays due on every tick, so the visit
countdowns of repair, transfer and the view fence run as before.

The WAL and repair ring gauges count sleeping partitions too, so
they are now sums of what each partition showed at its last visit.
A debug build tests every sleeping partition on every tick and
fails on one with work. A release build tests eight per tick,
counts them in partition_missed_wakeups and wakes them. The
simulator and its fuzzer can make every tick visit every
partition, and tests require the same executor schedule in both
modes, also across crashes, restarts and state transfers.

Three idle nodes with 30,000 partitions each: 350% -> 13% of one
core per node. A node visited 30,000 partitions on each of its 100
ticks per second, and partition_tick_visits now counts 50,000
visits per second.