Enforce that macrocompiler passes are done serially

[abejgonzalez]

If you look at the CI, there are multiple points where the HwachaRocketConfig fails during a MacroCompiler step with the following:

• java.lang.NoClassDefFoundError: mdf/macrolib/PortPolarity: https://circleci.com/gh/ucb-bar/chipyard/12736
• java.lang.ClassFormatError: Truncated class file: https://circleci.com/gh/ucb-bar/chipyard/13029

I haven't been able to replicate the error (yay non-determinism), but I theorize that this is a problem with 2 sbt runs on the same binary (e.g. MacroCompiler) running at the same time (since all our CI uses parallel make). sbt seems to only lock at the beginning when looking up things in ivy, so if MacroCompiler is getting compiled twice, it might clobber files?

This is a proposed solution by making the MacroCompiler serialized by having an order only dependency (https://www.gnu.org/software/make/manual/html_node/Prerequisite-Types.html) on the harness MacroCompiler pass.

Any thoughts on this?

[colinschmidt]

LGTM. I think we should see leave a low priority issue open to try to address this in a better way in the future.
For large designs with lots of potential memories (or when a true compiler is present) this non-overlap could cost a reasonable amount of time.

[abejgonzalez]

Added #396