sql,kv: power ColIndexJoin by COL_BATCH_RESPONSE scan format

The initial work on the KV projection pushdown will only utilize the newly-introduced COL_BATCH_RESPONSE scan format for simple scans (i.e. `TableReader`s). We should extend that also to index joins since we already have the vectorized `ColIndexJoin` (unlike the lookup joins that still go through the row-by-row engine).

The main contribution here would be to teach the `Streamer` to use the new scan format.

Also an important question that will need to be answered as part of this item is whether we squash `coldata.Batch`es from multiple requests into one or not. This is important since index joins read only a single row, so setting up the whole pushdown pipeline for each request separately can lead to significant overhead. One idea that was mentioned previously was to introduce a new request type (something like `BatchedScanRequest`) which would contain key spans that need to be scanned and can be returned as a single response. For index joins we don't care to know which particular input row resulted in which response row.

Jira issue: CRDB-23150