Optimize deserialization

As @fornwall notes:

> The generated deserializers track duplicate fields with a `HashSet`, which hashes each insertion and allocates storage as fields are recorded.

This could be avoided with a bitset: each field could get a known index between 0 and N, and N is known at compile time.

I am opening this issue so we can discuss ideas, tradeoffs, and quantify the wins.

Start by reviewing the generated deserializers and how their HashSet tracks duplicate fields. Compare the proposed compile-time field indexes and bitset approach, then quantify the tradeoffs and performance wins. The issue does not yet define a specific implementation or completion criteria, so those need agreement before work begins.

Written by the indexing model from the issue text.