storage: MVCCIncrementalIterator ignores intents below StartTime

**Describe the problem**

The MVCCIncrementalIterator ignores intents below the StartTime:

```
metaTimestamp := i.meta.Timestamp.ToTimestamp()
if i.startTime.Less(metaTimestamp) && metaTimestamp.LessEq(i.endTime) {
```

This behaviour is intentional, but it poses at least two problems:

## Potential Bug: ResolveIntent beneath closed timestamp can result in follower-reads skipping an intent they shouldn't skip

The following is an unverified hypothetical timeline:

1. Txn A writes k@ts1.
2. Txn B encounter k and pushes Txn A to ts2
3. Txn B resolves intent on k to ts2.
4. Application of resolve intent from (3) is delayed on some follower.
5. Range for k closes timestamp ts4.
6. Txn B has it's Write timestamp pushed to ts3.
7. Txn B sends a Refresh for `k (ts1, ts3]`. Since ts4 is closed, this can be served by the delayed follower from (4).
8. Refresh erroneous succeeds.

## Problem For VIR

Virtual Intent Resolution aims to reduce the impact of intents on readers and set the groundwork for non-blocking readers by allowing readers to "virtually resolve" intents in their local read snapshot without actually committing a write batch that resolves the intent. In this world, encountered intents will _often_ stay at their original timestamp on disk, meaning that Refresh and RefreshRange would erroneously skip them in the common case.

Jira issue: CRDB-66399

Epic CRDB-63288

Start at MVCCIncrementalIterator and inspect the intent timestamp condition, then trace how Refresh and RefreshRange consume its results. Done means the behavior is specified and the reported follower-read and VIR cases are covered without incorrectly skipping intents.

Written by the indexing model from the issue text.