Optimize Serde implementations for rosidl_runtime_rs types

[nnmm]

Particularly:

• The String Deserialize implementation first creates a Rust String, but it could potentially directly construct a rosidl_runtime_rs::String. Same for WString and the sequences types (which use a temporary Vec).
• The String Serialize implementation also creates a Rust String even though it wouldn't need to.

The sequence serialization implementations are already as efficient as they can be, I believe.

This is a tricky task that is best suited for someone who's quite familiar with the Serde data model, so I'm adding the needs-rust-expert label.