Report a shard served with fewer replicas than its table asks for

[bjmeetsfo]

Placement is worked out per request: a topology answer names whatever
servers could be found for each shard. When there are not enough -- too few
live ones, or none the separation rules will accept -- the answer is simply
shorter, and the caller uses it.

Nothing counted that. A table configured for three replicas can be served
with one indefinitely, and the only place it shows is a client noticing it
was handed fewer endpoints than it asked for. There is a readiness field
called under_replicated_repair_ready and it is a hardcoded true; nothing
measured the thing it is named after.

temporalstore_meta_placement_short_total counts the shards each answer could
not fill, and temporalstore_meta_placement_short is what the last answer was
short of. Counted where the answer is built, because placement happens per
request and no background round would otherwise notice.

A shard with no primary is out of service deliberately -- frozen, or its
owner not serving -- so it is not counted as short. It is stopped, not
under-replicated.

Test: one server, a table asking for three replicas, two shards. Both are
served with one replica and both are reported; a second answer moves the
total and leaves the gauge where it was.