Bound the number of impacts per block [LUCENE-9156]

Today we write all competitive impacts in skip lists. Maybe we should be more cautious: in the worst-case scenario, there could be as many impacts as postings in a block and this would not only waste space but also slow down queries due to the time it would take to decode all these impacts.

One idea would be to merge all impacts pairwise, taking the maximum term frequency and the minimum norm, until their number is under some hard-coded limit, e.g. 8.

---
Migrated from [LUCENE-9156](https://issues.apache.org/jira/browse/LUCENE-9156) by Adrien Grand (@jpountz), updated Jan 22 2020

Start with the migrated LUCENE-9156 proposal and the code that writes competitive impacts in skip lists. Determine how a hard-coded impact bound and pairwise merging would affect space use and query-time decoding. Done means the number of impacts per block is bounded without losing the stated maximum term frequency and minimum norm behavior.

Written by the indexing model from the issue text.