1) save memory by not using large slices when we only need small ones
  In particular, fix #1912
2) use a default size that corresponds to a capacity that append()
   actually uses (I believe a size of 800 gets rounded up by make)

e.g. in docker-dev with 1s precision 2days means 172800 points.
This is *before* maxdatapoints runtime consolidation,
so we may as well accommodate it with some special size classes
this makes sure our large slices won't be taken up by smaller (4k)
requests