This PR improves bloom filter performance.

The existing code computes the word and bit index using a straight-forward div. This is a problem because 1. the word location is a necessary dependency for the rest of the loop body, so this stalls the pipeline and 2. divs are very slow, obviously.

Thankfully, ClickHouse already has libdivide, so converting this to a mul using the reciprocal value is easy. While doing this, also removed one mul in the body and turned it into an add.

Lastly, ensured that word_size is marked as a constant, so the compiler can always emit shifts and masks for the word_size divisions and modulo ops.

All of this together makes the primary bottleneck in the BloomFilter functions CityHash, as it should be.
We have inspected this as carefully as possible to ensure mathematical equivalence and have tested it on existing data with no issues.

In our very large production cluster, we saw index time per byte in merges drop by ~45%:

Query:

SELECT
  toStartOfInterval (event_time, interval 5 minute) time,
  sum(
    ProfileEvents[
      'MergeTreeDataWriterSkipIndicesCalculationMicroseconds'
    ]
  ) / sum(size_in_bytes) index_us_per_byte
FROM
  clusterAllReplicas('{cluster}', system.part_log)
WHERE
  table = 'main'
  and event_date > now() - interval 14 day
  and $__timeFilter(event_time)
  and event_type = 'MergeParts'
  and size_in_bytes > 0
group by
  1
order by
  1 asc

Changelog category (leave one):

• Performance Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

• Performance improvements to all bloom filter types.