Fix ref-count for multiple stores to the same pubkey in a slot, fixes zero lamport purge detection

[carllin]

Problem

Storing to the same key multiple times in a slot only keeps one entry alive for that (pubkey, slot) pair in the AccountsIndex, but increments the ref count multiple times. This means the zero-lamport purge logic will not accurately detect if an account can be purged

Summary of Changes

Only increment ref count once per store per pubkey/slot.

Fixes #

[codecov]

Codecov Report

Merging #12462 into master will increase coverage by 0.0%.
The diff coverage is 100.0%.

@@           Coverage Diff           @@
##           master   #12462   +/-   ##
=======================================
  Coverage    82.0%    82.0%           
=======================================
  Files         354      354           
  Lines       82719    82732   +13     
=======================================
+ Hits        67896    67913   +17     
+ Misses      14823    14819    -4     

[ryoqun]

how about reduce()-ing here like this, which will be done in parallel by rayon?:

let pubkeys: HashSet<(Slot, Pubkey)> = 
   ...
   .reduce(|| HashSet<_>::new(), |mut reduced, pubkeys| reduced.extend(pubkeys)

so that we can remove the need to construct cleaned_slot_keys later?

[carllin]

updated!

[ryoqun]

LGTM with nits, awesome bug fixing. I guess you found this by carefully reading the code to fight with the snapshot bloat issue at #12194 .

I guess this bug doesn't manifest as clear errors. Only this causes is dangling appendvecs, which doesn't reduce the snapshot size as expected.

Pull request has been modified.

[carllin]

@ryoqun I actually found it while writing those snapshot hash mismatch tests :)