[multistage][bug] block splitter estimation is way off

when we send data over the mailboxes we are estimating the data size and cut the inbound messges into chunks. however

```
// Use estimated row size, this estimate is not accurate and is used to estimate numRowsPerChunk only.
int estimatedRowSizeInBytes = block.getDataSchema().getColumnNames().length * MEDIAN_COLUMN_SIZE_BYTES;
int numRowsPerChunk = maxBlockSize / estimatedRowSizeInBytes;
while (currentRow < totalNumRows) {
List chunk = allRows.subList(currentRow, Math.min(currentRow + numRowsPerChunk, allRows.size()));
```

this `estimatedRowSizeInBytes` can be WAY OFF when there's high-cardinality string/bytes column, and those columns are super large.

simple solution is to use the first row to estimate the size of the row when there's variable length columns found, but
- there's no easy way to tell cardinality
- it is expensive to compute a row size of `Object[]` which needs to loop through everything.

Locate the multistage block splitter and the estimation using getDataSchema(), MEDIAN_COLUMN_SIZE_BYTES, and maxBlockSize. Reproduce the problem with a block containing a very large variable-length string or bytes column, then inspect existing splitter coverage and add a regression case. Done means chunk sizing is no longer badly underestimated while preserving the mailbox size limit.

Written by the indexing model from the issue text.