Why?

Currently Rust uses a boolean write_ref_info/read_ref_info parameter which only distinguishes between "write ref flag" vs "don't write ref flag". This doesn't support:

• Nullable types without ref tracking (like Option<T> where T is not a shared ref)
• Explicit control over shared reference tracking per field via #[fory(ref=false)] attribute
• Alignment with Go's RefMode design which has three modes: None, NullOnly, Tracking

What does this PR do?

1. Add RefMode enum to replace boolean write_ref_info/read_ref_info

#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
#[repr(u8)]
pub enum RefMode {
    #[default]
    None = 0,      // No ref flag written/read (primitives, non-nullable types)
    NullOnly = 1,  // Write NULL(-3) or NOT_NULL(-1), no ref tracking
    Tracking = 2,  // Full ref tracking with REF(-2) and REF_VALUE(0) flags
}

2. Update Serializer trait signature

fn fory_write(&self, context: &mut WriteContext, ref_mode: RefMode, write_type_info: bool, has_generics: bool) -> Result<(), Error>
fn fory_read(context: &mut ReadContext, ref_mode: RefMode, read_type_info: bool) -> Result<Self, Error>

3. Per-field RefMode determination in derive macro

• Type classification: primitive → None, Option → NullOnly, Rc/Arc/Weak → Tracking
• Field meta attributes respected: #[fory(ref=false)] to disable tracking, #[fory(nullable)] for null support
• All field types (including trait objects like Box<dyn Any>, Rc<dyn Trait>, etc.) now respect field attributes

4. Add FieldRefMode enum with ToTokens impl for cleaner macro code generation

pub(crate) enum FieldRefMode {
    None,
    NullOnly,
    Tracking,
}

impl ToTokens for FieldRefMode {
    fn to_tokens(&self, tokens: &mut proc_macro2::TokenStream) {
        // Generates fory_core::RefMode::None/NullOnly/Tracking
    }
}

5. Add determine_field_ref_mode(field) helper

Combines type classification with field meta attributes to determine the correct RefMode for each field, respecting #[fory(ref=false)] and #[fory(nullable)] annotations.

Related issues

Closes #3097

Does this PR introduce any user-facing change?

• Does this PR introduce any public API change?
  • Serializer::fory_write and Serializer::fory_read now take RefMode instead of bool
  • New RefMode enum exported from fory_core
• Does this PR introduce any binary protocol compatibility change?
  • No, the wire format is unchanged

Benchmark

No performance regression expected - the RefMode enum comparison (ref_mode != RefMode::None) compiles to the same machine code as the previous boolean comparison.