This fixes #654.

There are only two important changes.

1) The first one fixes missing dependencies:

```diff
-pkgDependencies = fmap (fmap PD.dependencies) . readPackageDataFile
+pkgDependencies = fmap (fmap (map pkgName . packageDependencies)) . readCabalFile
```

Here `PD.dependencies` returned versioned package names, e.g. `ghc-8.7`, which then failed to match with non-versioned package names such as `ghc` in `contextDependencies`. Switching from `PD.dependencies` to `packageDependencies` fixes this.

2) I clearly remember that we didn't have this bug before, so I added some tests for our package dependency infrastructure to prevent such regressions in future:

```haskell
testDependencies :: Action ()
testDependencies = do
    putBuild "==== pkgDependencies"
    depLists <- mapM (pkgDependencies . vanillaContext Stage1) ghcPackages
    test $ and [ deps == sort deps | Just deps <- depLists ]
    putBuild "==== Dependencies of the 'ghc-bin' binary"
    ghcDeps <- pkgDependencies (vanillaContext Stage1 ghc)
    test $ isJust ghcDeps
    test $ pkgName compiler `elem` fromJust ghcDeps
    stage0Deps <- contextDependencies (vanillaContext Stage0 ghc)
    stage1Deps <- contextDependencies (vanillaContext Stage1 ghc)
    stage2Deps <- contextDependencies (vanillaContext Stage2 ghc)
    test $ vanillaContext Stage0 compiler `notElem` stage1Deps
    test $ vanillaContext Stage1 compiler `elem`    stage1Deps
    test $ vanillaContext Stage2 compiler `notElem` stage1Deps
    test $ stage1Deps /= stage0Deps
    test $ stage1Deps == stage2Deps
```

Everything else are cosmetic changes, fixing minor issues in comments, and adding TODOs. To figure out the failure in #654 I had to read some code I didn't write and my hands were automatically fixing some style inconsistencies with the rest of the Hadrian code base. (I'd like to emphasise that I make no judgement about which style is better, it's merely an attempt to make the code base look more homogeneous, which I think is useful.)