Fix JSON parsing bug for modeled empty structs

[jdisanti]

Motivation and Context

This will fix aws-sdk-rust#212.

Testing

• Reproduced the issue with the following:

#[tokio::main]
async fn main() {
    tracing_subscriber::fmt::init();

    let resp = http::Response::builder()
        .status(200)
        .body(
            r#"
            {"scheduleActions":[{"actionName":"REDACTED","scheduleActionStartSettings":
            {"immediateModeScheduleActionStartSettings":{"time":"2021-08-29T21:54:42.651"}},
            "scheduleActionSettings":{"inputSwitchSettings":
            {"inputAttachmentNameReference":"REDACTED","urlPath":[]}}}]}
            "#,
        )
        .unwrap();

    let _schedule = DescribeSchedule::new()
        .parse_loaded(&resp.map(Bytes::from))
        .unwrap();
}

• Also reproduced the issue in the JsonParserGeneratorTest.
• Added rest-json-extras test case.
• Verified all work correctly after code fix.

Checklist

• I have updated the CHANGELOG

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[jdisanti]

Diff of generated code from these changes

[rcoh]

if we actually put the members.isEmpty check inside of objectKeyLoop, I think we can remove the special casing and the other changes around consuming vs. not consuming the start element? It also makes it a little clearer and more consistent IMO.

Am I correct in saying that if we fully removed the members.isEmpty() branch the code would be correct (but maybe it would fail a clippy check?)

[jdisanti]

if we actually put the members.isEmpty check inside of objectKeyLoop, I think we can remove the special casing and the other changes around consuming vs. not consuming the start element? It also makes it a little clearer and more consistent IMO.

I think overall, it's simpler, but it does require the introduction of a skip_to_end variation of skip_value that assumes the start object/array has already been consumed.

Am I correct in saying that if we fully removed the members.isEmpty() branch the code would be correct (but maybe it would fail a clippy check?)

Yeah, that's correct.

[jdisanti]

There's no codegen diff now for the smoketest services since the only difference is for modeled empty structs, and none of them have any. Here's an example of generated code for a modeled empty struct though:

pub fn deser_structure_empty_struct<'a, I>(
    tokens: &mut std::iter::Peekable<I>,
) -> Result<Option<crate::model::EmptyStruct>, smithy_json::deserialize::Error>
where
    I: Iterator<
        Item = Result<smithy_json::deserialize::Token<'a>, smithy_json::deserialize::Error>,
    >,
{
    match tokens.next().transpose()? {
        Some(smithy_json::deserialize::Token::ValueNull { .. }) => Ok(None),
        Some(smithy_json::deserialize::Token::StartObject { .. }) => {
            #[allow(unused_mut)]
            let mut builder = crate::model::EmptyStruct::builder();
            smithy_json::deserialize::token::skip_to_end(tokens)?;
            Ok(Some(builder.build()))
        }
        _ => Err(smithy_json::deserialize::Error::custom(
            "expected start object or null",
        )),
    }
}