retains one node-instance per pubkey (#17187)

crds table retains up to 32 node-instance values per each pubkey. This
is so because if there are multiple running instances of the same node,
then we want gossip to propagate node-instance values associated with
both instances, therefore the corresponding label/key includes the
randomly generated token in addition to the pubkey:
https://github.com/solana-labs/solana/blob/9c42a89a4/core/src/crds_value.rs#L448
https://github.com/solana-labs/solana/pull/14037

As a result, the number of such values per pubkey are effectively
unbounded, requiring custom mitigations implemented in:
https://github.com/solana-labs/solana/pull/14467
but still taking redundant extra memory and bandwidth.

This commit instead retains only one node-instance per pubkey by
extending crds values override logic. If a crds value is of type
node-instance, it will always override an existing one with the same key
if it has more recent starting timestamp (not wallclock). As a result,
gossip will always propagate the node-instance with more recent
timestamp. Since the check_duplicate logic will stop the node with older
timestamp, this change should preserve existing functionality.

core/src/cluster_info.rs

@@ -454,10 +454,10 @@ fn retain_staked(values: &mut Vec<CrdsValue>, stakes: &HashMap<Pubkey, u64>) {
             // Otherwise unstaked voting nodes will show up with no version in
             // the various dashboards.
             CrdsData::Version(_) => true,
+            CrdsData::NodeInstance(_) => true,
             CrdsData::LowestSlot(_, _)
             | CrdsData::AccountsHashes(_)
             | CrdsData::LegacyVersion(_)
-            | CrdsData::NodeInstance(_)
             | CrdsData::DuplicateShred(_, _) => {
                 let stake = stakes.get(&value.pubkey()).copied();
                 stake.unwrap_or_default() >= MIN_STAKE_FOR_GOSSIP