Use YAML syntax for remote git location identification.

As suggested by @greggwebs. The syntax proposed yesterday has the
advantage of concision, but using YAML syntax for the different
components adds marginal verbosity. As it is, the previous syntax made
extensibility problematic, whereas that's exactly where YAML
shines (what if we need an extra field to identify a particular git
commit?). Further, using YAML means people are free to include their own
metadata about package locations, without disturbing Stack.

Here's an example:

- location:
    git: https://github.com/kolmodin/binary
    commit: 8debedd3fcb6525ac0d7de2dd49217dce2abc0d9

src/Stack/Types/Config.hs

@@ -234,17 +234,16 @@ instance ToJSON PackageLocation where
     toJSON (PLHttpTarball t) = toJSON t
     toJSON (PLGit x y) = toJSON $ T.unwords ["git", x, y]
 instance FromJSON PackageLocation where
-    parseJSON = withText "PackageLocation" $ \t ->
-        http t <|> git t <|> pure (PLFilePath $ T.unpack t)
+    parseJSON v = git v <|> withText "PackageLocation" (\t -> http t <|> file t) v
       where
+        file t = pure $ PLFilePath $ T.unpack t
         http t =
             case parseUrl $ T.unpack t of
                 Left _ -> mzero
                 Right _ -> return $ PLHttpTarball t
-        git t =
-            case T.words t of
-                ["git", x, y] -> return $ PLGit x y
-                _ -> mzero
+        git = withObject "PackageGitLocation" $ \o -> PLGit
+            <$> o .: "git"
+            <*> o .: "commit"
 
 -- | A project is a collection of packages. We can have multiple stack.yaml
 -- files, but only one of them may contain project information.