Fix number deserialization with arbitrary_precision

[bkchr]

When we have a struct that uses #[serde(flatten)] and the feature
arbitrary_precision is enabled, the deserialization. The problem was
that the arbitrary number was forwarded as a "map" to the visitor and in
a later step this map failed as a number was expected.

Fixes: #505

Requires: serde-rs/serde#1679

[bkchr]

I removed this test + the extra added function from_string_unchecked. In my opinion this test does not makes that much sense at all. Even if we would have parsed the 1e999 correctly into the Number, there is not way to get this value out of the object.

[WiSaGaN]

The patch seems still failing the test, which would have been passed without "arbitrary_precision" feature.

use serde::{Deserialize};

#[derive(Deserialize)]
#[serde(tag = "tag")]
#[derive(Clone, Debug, PartialEq)]
pub enum E1 {
    A{v: f64},
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn serde_1() {
        let incoming_str = r#"{"tag":"A","v":0}"#;
        let expected_incoming = E1::A{ v: 0.0 };
        let incoming = serde_json::from_str(&incoming_str).expect("cannot deserialize incoming_str");
        assert_eq!(expected_incoming, incoming);
    }
}

"invalid type: unexpected i128, expected f64", line: 0, column: 0

[dtolnay]

Hi @bkchr, thanks for the PR but I can't accept this because it breaks the following:

// [dependencies]
// serde_json = { version = "1.0", features = ["arbitrary_precision"] }

fn main() {
    println!("{}", serde_json::from_str::<serde_json::Value>("1e999").unwrap());
}

This is what is tested by the test you removed. The point of arbitrary_precision is to allow deserializing and round-tripping arbitrarily large numbers without losing precision.