Add some more tests for config validation, and fix some gaps (#274)

I set these tests up in #270, but I realized there's a lot more we could
test unrelated to that PR. In this commit I add some more tests. They're
not really exhaustive yet, but they did catch a few bugs, which I fixed:
- we weren't validating the package-name if you do set it (only if we
guess it)
- If you omitted `generated`, you would try to write generated code to
the directory containing `genqlient.yaml`, which makes no sense; now we
default to `generated.go`. In the real world it's probably good to set
explicitly, but it's actually very convenient in tests that we don't
have to, and maybe in small projects too.

docs/genqlient.yaml

@@ -26,7 +26,7 @@ operations:
 - "pkg/*.go"
 
 # The filename to which to write the generated code, relative to
-# genqlient.yaml.
+# genqlient.yaml. Default: generated.go.
 generated: generated/genqlient.go
 
 # The package name for the output code; defaults to the directory name of
@@ -77,7 +77,6 @@ context_type: context.Context
 # without making a query.
 client_getter: "github.com/you/yourpkg.GetClient"
 
-
 # If set, fields with a struct type will default to having
 # the "pointer: true, omitempty: true" flag.
 #