mGCA: Support directly represented negated literals

[BoxyUwU]

Follow up of #150618 and #150699. We currently support directly represented literals as a hir::ConstArgKind::Literal but we don't support negated literals as such a representation:

#![feature(adt_const_params, min_generic_const_args)]

use std::marker::ConstParamTy;

#[derive(Eq, PartialEq, ConstParamTy)]
struct Foo {
    field: isize
}

fn foo<const F: Foo>() {}

fn main() {
    // works
    foo::<{ Foo { field: 1} }>();
    // needs an anon const
    foo::<{ Foo { field: -1 } }>();
    // works
    foo::<{ Foo { field: const { -1 } } }>();
}

error: complex const arguments must be placed inside of a `const` block
  --> src/main.rs:16:26
   |
16 |     foo::<{ Foo { field: -1 } }>();
   |                          ^^

Not certain of the right way of implementing this, maybe extending ConstArgKind::Literal with a flag for whether it was negated or not 🤔 Then it should just be a matter of updating lower_anon_const_to_const_arg_direct with a case for -1 as well as threading the bool into the input of lit_to_const in lower_const_arg_literal

Once this and #152001 have been done we should remove the special casing in parsing under mGCA for literals being parsed as anon consts instead of direct const args.

[khyperia]

@rustbot claim