Use an approximate percentile algorithm by default

[nblumhardt]

Seq's percentile() aggregation is exact, which, while useful in many cases, is resource-intensive and parallelizes poorly (Seq 2023.1 can parallelize percentile() calculations that group by time(..) but otherwise cannot parallelize percentiles).

For most purposes, using an approximate percentile algorithm by default, and supporting exact calculations through a percentile_exact or similar, seems to make more sense.