RAM usage

It's annoying to have to slow a big build (ie tensorflow) right down with
`--jobs 1`
just for the odd job that is memory hungry, because Bazel isn't being smart about allocating work. Why can't the job scheduler just check the current RAM usage and only add a job if it's under the utilization limit? Then _if_ there's room it can add jobs until all cores are busy.

The issue does not name files, tests, or an entry point. Start by locating Bazel's job scheduler and its existing resource-limit behavior, then define how RAM utilization should affect job admission and validate it against a large build such as TensorFlow.

Written by the indexing model from the issue text.