release-21.1: kvserver: use ClearRawRange to truncate very large Raft logs

[erikgrinaker]

Backport 1/1 commits from #75793.

/cc @cockroachdb/release

Release justification: possible node crash loop.

---

Raft log truncation was done using point deletes in a single Pebble
batch. If the number of entries to truncate was very large, this could
result in overflowing the Pebble batch, causing the node to panic. This
has been seen to happen e.g. when the snapshot rate was set very low,
effectively stalling snapshot transfers which in turn held up log
truncations for extended periods of time.

This patch uses a Pebble range tombstone if the number of entries to
truncate is very large (>100k). In most common cases, point deletes are
still used, to avoid writing too many range tombstones to Pebble.

Release note (bug fix): Fixed a bug which could cause nodes to crash
when truncating abnormally large Raft logs.

[cockroach-teamcity]

This change is 

[tbg]

I know the pebble folks feel pretty good about rangedels in 21.2+, but is there something on 21.1 that makes them pause? I assume not, and also the tombstones are only used when you'd likely have a really bad time without them anyway.

[erikgrinaker]

Yeah, I thought 21.1 was probably recent enough that it's fine, but worth checking.

@cockroachdb/storage Any significant issues with range tombstones in 21.1?

[itsbilal]

@cockroachdb/storage Any significant issues with range tombstones in 21.1?

None, relative to 21.2 at least. This change should be good on 21.1 as well. There's cockroachdb/pebble#1070 but that covers both versions and is a more general issue.

[erikgrinaker]

Thanks @itsbilal!