Lazily allocate DocAndScoreAccBuffer in MaxScoreBulkScorer

[romseygeek]

This buffer is only used in certain code paths, so allocation in the constructor adds unnecessary costs.

[iverase]

LGTM

[iprithv]

thanks for the fix, @romseygeek.
Apologies.. this regression was introduced in #15971, where I replaced the cardinality() call with growNoCopy(INNER_WINDOW_SIZE) directly in the MaxScoreBulkScorer constructor to avoid a redundant bitset pass. while that helped the hot multi-essential-clause path, it caused unnecessary eager allocation of the large DocAndScoreAccBuffer arrays in many other cases (especially boolean queries with expensive two-phase subclauses, which affected some percolator queries).
I see that you have restored lazy allocation for the buffer while preserving the performance benefit. thanks again...Tagging #15971 for traceability.

[romseygeek]

No worries @iprithv, this didn't get picked up in any of the nightly benchmarks and it only really effects some corner cases (like elasticsearch's percolator) that ought to be disabling bulk scoring in any case. I think this is probably still worth doing to make things slightly more efficient when we use the other code paths, but it isn't a serious regression. I'll put it into 10.6