Shrink CI matrix

[lihaoyi]

• We probably only need to each on each Java and Scala version once, rather than on the full cross matrix of all combinations.
• We probably don't need to test every PR with every single Scala version. For now, for each Scala major version I picked three minor versions: highest, lowest, and one in the middle.
• Replaced __.publishLocal with __.compile, which should hopefully catch most of the same bugs but be much cheaper not creating jars
• I suspect the unitTest and integrationTest helper tasks were not doing the right thing - they were taking far longer than they should have been taking - but I couldn't figure out why. For now I just ditched them and use Mill queries to select the tests I want manually
• Turned on Mill parallelism
• We drop CI for Java 8, in line with the rest of the com-lihaoyi ecosystem

Hopefully that will give us enough confidence that things mostly work, and if anything emerges that's super version specific we can fix it ad-hoc without needing to wait 40+ hours of CPU time for every PR to be validated