Pulsar 2.9.2: messages are in backlog, but not dispatched to consumers

**Describe the bug**
I have Pulsar 2.9.2 installed on virtual servers, 9 brokers, 9 bookeepers.

We have plenty of topics in this cluster, all of them are consumed by the same application (our internal queue system).

For several topics we can see that there are some time some messages are not dispatched to consumers.
Here what we can see in internal-stats:

```
"subscriptions": {
"": {
"blockedSubscriptionOnUnackedMsgs": false,
"isReplicated": false,
"msgRateOut": 7.916492400494186,
"msgThroughputOut": 1755.328026537821,
"msgRateRedeliver": 0,
"msgRateExpired": 0,
"msgBacklog": 3,
"msgDelayed": 2,
"unackedMessages": 0,
"type": "Shared",
"activeConsumerName": "",
...
```
[stat.txt](https://github.com/apache/pulsar/files/9079227/stat.txt)

internal-stats:
```
...
"cursors": {
"": {
"markDeletePosition": "23451824:5393",
"readPosition": "23497178:23031",
"waitingReadOp": true,
"pendingReadOps": 0,
"messagesConsumedCounter": 502849,
"cursorLedger": 23495692,
"cursorLedgerLastEntry": 3950,
"individuallyDeletedMessages": "[(23451824:5394..23451824:49999],(23458318:-1..23458318:29820],(23466518:-1..23466518:49999],(23474137:-1..23474137:49999],(23478437:-1..23478437:49999],(23481746:-1..23481746:49999],(23485116:-1..23485116:49999],(23489019:-1..23489019:49999],(23493059:-1..23493059:49999],(23497178:-1..23497178:23028]]",
"lastLedgerWitchTimestamp": "",
"state": "Open",
"numberOfEntriesSinceFirstNotAckedMessage": 447459,
"totalNonContiguousDeletedMessagesRange": 10,
"properties": {}
}
...
```
[internal-stat.txt](https://github.com/apache/pulsar/files/9079229/internal-stat.txt)

As you can see - there are several messages that are not acked. But in our logs we see that they were not even delivered to consumers! If we perform topic unload - then they will be redelivered and consumed successfully.

**To Reproduce**
Do not know exactly the way of reproduction

Software versions:

- pulsar 2.9.2
- client: golang

**Expected behavior**
Messages should be delivered normally

Start by examining the attached internal-stats and stat.txt output, focusing on the subscription backlog, cursor positions, and pending read state. Compare normal consumer delivery with what happens after a topic unload; done means identifying why messages remain in backlog without delivery and confirming they are dispatched without requiring an unload.

Written by the indexing model from the issue text.