29129: backport-2.1: storage: Up-replicate before removing replicas from dead stores r=a-robinson a=a-robinson

Backport 1/1 commits from #28875.

/cc @cockroachdb/release

---

Fixes #25392, by preventing the following series of events:

1. Node x dies
2. We remove node x's replica of some range
3. Node y dies before we up-replicate, leaving the range unavailable (1
   out of 2 replicas dead)
4. Node x comes back online. It can't help the situation, because its
   replica was officially removed from the range.

Instead, we now up-replicate first before removing node x's replica.

Release note: None


29177: release-2.1: server: add a separate sampling loop for Go mem stats r=tschottdorf a=petermattis

Backport 5/5 commits from #29061.

/cc @cockroachdb/release

---

`runtime.ReadMemStats` will block waiting for any running GC cycle to
finish. On a large heap this can take upwards of 10s which is in excess of
the default frequency that we sample runtime statistics (10s).

Fixes #27775

Release note: none


Co-authored-by: Alex Robinson <[email protected]>
Co-authored-by: Peter Mattis <[email protected]>