cache: readShard: use a separate mutex

[RaduBerinde]

We use a separate readShard mutex to to protect the readMap and
each entry's ref count. This removes the possibility of a couple of
cornercases, making the code much cleaner. We also now avoid
write-locking the entire cache shard when we start a read.

This will also potentially allow moving read buffer management into
readShard in the future (and having separate readShard sets per
store).

[cockroach-teamcity]

This change is 

[sumeerbhola]

Reviewed 4 of 4 files at r1, all commit messages.
Reviewable status: all files reviewed, 3 unresolved discussions (waiting on @RaduBerinde)

---

internal/cache/refcnt_tracing.go line 54 at r1 (raw file):

}

func (v *refcnt) value() int32 {

this isn't needed either, yes?

---

internal/cache/read_shard.go line 51 at r1 (raw file):

//     creating a readEntry that is no longer needed because someone else has
//     done the read since the miss and inserted into the cache. By making the
//     readShard use shard.mu, such a race is avoided. A side benefit is that

This comment is a bit stale.

---

internal/cache/read_shard.go line 59 at r1 (raw file):

//     is the possibility that the compaction reader and a user-facing iterator
//     reader will do duplicate reads, but we accept that deficiency.
//     TODO(radu): consider using a separate mutex for the readShard; we are

Isn't this done by this PR?