fix(schema): remove AddNullable from draft2020_12 settings

[mwotton]

Problem

The schema_for_type() function uses AddNullable::default() as a transform when generating JSON Schema 2020-12 schemas. However, the nullable keyword is an OpenAPI 3.0 extension, not part of JSON Schema 2020-12. This causes validation failures with strict JSON Schema validators.

Solution

Remove the AddNullable transform from the draft2020_12 settings. JSON Schema 2020-12 represents nullable types using:

• {"type": ["string", "null"]} (type array with null)
• {"anyOf": [{"type": "string"}, {"type": "null"}]}

Before

An optional string field produced:

{
  "nullable": true,
  "type": "string"
}

After

An optional string field produces:

{
  "type": ["string", "null"]
}

Fixes #663

[alexhancock]

@mwotton Thanks for the fix. test_complex_schema needs a look

[mwotton]

Added a second follow-up commit (70011f9) after the test_complex_schema fix because another test had the same stale nullable assumption.

test_tool_macros::test_optional_field_schema_generation_via_macro was still asserting OpenAPI-style nullable: true and explicitly rejecting type: ["string", "null"], which conflicts with Draft 2020-12 output from schema_for_type().

This was not visible in the original CI failure because cargo test --all-features failed fast on test_complex_schema first. The second commit aligns that macro test with Draft 2020-12 behavior.

[DaleSeo]

Hey @mwotton, thanks for the fix! The code change looks good.

The CI failures aren't related to your changes. All five failing checks hit the same compilation error:

error[E0004]: non-exhaustive patterns: `Some(Protocol::Tlsv13)` not covered  
--> native-tls-0.2.17/src/imp/openssl.rs:61:22

This happened because your branch is based on an older main that still used reqwest 0.12, which depends on native-tls. Since then, PR #669 upgraded to reqwest 0.13, which uses rustls by default and doesn't pull in native-tls, so the main CI is green now.

To fix the CI, please rebase your branch onto the latest main. I think that should resolve all five failures. The code changes themselves look correct.