Filtered read performance optimizations

There are a few bottlenecks in the planning phase of filtered reads that could probably be optimized. These are most noticeable when reading small amounts of data from a large dataset (e.g. applying a btree index) when all the data is in memory.

• RowIdTreeMap::insert_range seems kind of slow. This makes RowIdTreeMap::from(range) slow which means RowIdSequence::mask_to_offset_ranges is slow which means apply_index_result in FilteredReadStream::create_plan is slow.
• The apply_index_result step is run once for each fragment, in serial order. It is not parallelized because that would mess up range_offset and filtered_range_offset. However, we could probably calculate just those offsets first (in a vec) and then parallelize the apply_index_result step

Start by reading RowIdTreeMap::insert_range and tracing its use through RowIdTreeMap::from, RowIdSequence::mask_to_offset_ranges, and FilteredReadStream::create_plan. Examine how range_offset and filtered_range_offset are calculated before changing the serial apply_index_result work. Done means filtered-read planning is measurably faster while preserving correct offsets and results.