Arguably, emptyVariant is more like Void than (), given that it is
actually not possible to obtain its value via fromVariant and if it
were, a program would explode due to undefined.

In JSON there is not really a way to have a void value, a JSON value is
always *something*, so it makes no sense to have a Variant '[] parser
that succeeds in any condition. More importantly, the FromJSON instance
for Variant (t ': ts) actually *requires* the instance for Variant '[]
to fail! If parseJSON is called for Variant '[] from that instance, it
means that the value didn't match any of the types the Variant accepts.

This causes the bug that, if parseJSON :: Parser () succeeds for a JSON
value, we can obtain an arbitrary Variant that would fail to evaluate,
as it'd be constructed via extendVariant emptyVariant. For aeson < 1.6
this is a pretty obscure bug, since FromJSON () only accepts the empty
array:

    >>> decode "[]" :: Maybe (Variant '[Int])
    Just *** Exception: Prelude.undefined

Since aeson 2.0, FromJSON () [accepts] all JSON values, meaning that any
failed parse to Variant results in a program crash!

Tagged Variants also exhibit this problem, of course. The bug is fixed
by making the FromJSON (Variant '[]) instance fail always via parseFail.

[accepts]: haskell/aeson@677daf0

This adds a bit more ifdef zoo to be compatible with the breaking
change in aeson. Tested manually against aeson 2.0.3.0 and 1.5.6.0.

Might want to update the CI as well, but I don’t have the time to
touch CircleCI yaml.

Co-authored-by: Lukas Epple <[email protected]>

Support for * is disabled in GHC 9.2 by default and causes a lot of
annoying warnings in 9.0.

Tested against 8.10.7, 9.0.2 and 9.2.4. The base constraint seems to be
correct, as Data.Kind was added in base 4.9.

Co-authored-by: Lukas Epple <[email protected]>

This seems to be required now to compile the Variant code.

The previous show instance would add another layer of quoting for each
nesting:

```
ghci> show (rcons (#foo := "bar") rnil)
"[(\"foo\",\"\\\"bar\\\"\")]"
```

Instead, what we want is to display a nicely readable variant of the
record, using the infix field syntax for both label/value pairs and
full records:

```
ghci> show (rcons (#hi := (rcons (#lea := "hi") rnil)) (rcons (#foo := "bar") rnil ))
"[#foo := \"bar\",#hi := [(#lea := \"hi\")]]"

ghci> show (#hi := "lea")
"#hi := \"lea\""
```

That’s better!

Note that we can’t have a roundtripping `Read` instance anyway, so we
might as well have `Show` be readable.

Fixes agrafix#36