Apply filters to `RecordBatch` instead of indices in nested loop join

### Is your feature request related to a problem or challenge?

> I think batch coalescer won't make this faster as this is buffering everything in memory anyway.
>
> The main idea would be to apply the filters iteratively to the incoming `RecordBatch` instead of the indices, so we have to change the API / implementation a bit more.

@Dandandan mentions in #16443 that we apply filters to the `RecordBatch` when processing incoming batches in nested loop join.

### Describe the solution you'd like

_No response_

### Describe alternatives you've considered

_No response_

### Additional context

_No response_

Start by tracing how nested loop join processes incoming RecordBatch values and where filters are currently applied to indices; review the discussion in #16443. Done means defining the required API and implementation changes so filters are applied iteratively to the incoming RecordBatch instead.

Written by the indexing model from the issue text.