Some targets are tagged manual for other reasons and manual really just
means "don't build this in wildcards", which is a human-focused
distinction. For build bots, we can now use `bazel query` (which
ignores "manual") piped to `bazel test` to build and test all targets
in parallel (more important as we get more tests).

Tested:
Locally ran
```shell
bazel query //...   @llvm-project//... \
  | xargs bazel test \
    --config=rbe \
    --test_output=errors \
    --test_tag_filters=-nobuildkite \
    --build_tag_filters=-nobuildkite
```