Call the targets that already define these commands, so that the
workflow and a developer's machine run the same thing rather than two
copies of it that drift apart.

• Replace the dependency, test, archive and packaging commands with
  the targets that hold them
• Leave the certificate, notarization and upload steps alone, being
  the runner's business and not runnable locally

The targets these steps call came from #590, now merged, so this is a
single commit against master.

Where CI attaches to the target graph

The same picture as #590, which explains the targets themselves. What
matters here is which nodes the workflow now enters at.

Before:

flowchart TD
  pb["pre-build<br>builds deps inline"] --> gss["git-submodule-sync"]
  ds["dmg-signed<br>packages inline"] --> ar["archive"]
  bp["build-project"] --> ar
  app["app"] --> ar
  dmg["dmg"] --> app
  run["run"] --> build["build"]
  ut["unit-test"]
  uit["ui-test"]

After, with the three targets #590 adds shaded:

flowchart TD
  pb["pre-build"] --> gss["git-submodule-sync"]
  pb --> deps["deps"]
  ds["dmg-signed"] --> ar["archive"]
  ds -. "via make" .-> ps["package-signed"]
  bp["build-project"] --> ar
  app["app"] --> ar
  dmg["dmg"] --> app
  run["run"] --> build["build"]
  ut["unit-test"]
  uit["ui-test"]
  at["all-tests"]
  classDef new fill:#E1F5EE,stroke:#0F6E56,color:#04342C
  class deps,ps,at new

This PR points five steps at deps, unit-test, all-tests,
archive and package-signed. None of them is an aggregate: CI enters
low in the graph every time, while a person keeps using the targets
above them, pre-build, test and dmg-signed. That is why deps
and package-signed had to be split out of their parents at all. CI
has checked the submodules out through its own step and builds the
archive in another, so a step calling pre-build would redo the
checkout and a step calling dmg-signed would archive a second time
per architecture.

The duplicate test run is kept, deliberately

CI's "Run tests" step passes no -only-testing, and the scheme tests
every target, so it runs GitXTests a second time after the "Run unit
tests" step has already run them. This change keeps that exactly as it
is, calling make all-tests, which is the target written to match it.

make ui-test narrows the run to GitXUITests, which is all the
screenshots need, and would drop the repeat. That is a change to what
CI does rather than to how it says it, so it is left alone here and
marked with a TODO beside the step. Say the word and it is a one-word
edit.

Test plan:

• This PR's own CI run is green on both architectures, with make deps, make unit-test ARCH=arm64 and make archive running on the
  runners
• make all-tests ARCH=arm64 RESULT_BUNDLE=<path> is green locally and
  writes the bundle, running the unit suites and GitXScreenshotTests
• make -n for every converted step reproduces the command that step
  runs today, with the same paths
• The screenshot and packaging steps need signing secrets, so they do
  not run on a fork PR and are verified by reading and by make -n
  only