[Benchmark] Validate phf dispatch threshold with end-to-end XML decoding #57

Context

PR #56 adds opt-in Rust --feature phf dispatch and a Criterion lookup benchmark. The benchmark and hardware-counter data in docs/benchmarks/rust-phf-dispatch.md suggest a crossover around 835 element names, but dispatch alone does not measure full XML decoding. The current compiler does not automatically choose phf at a schema-size threshold.

Investigate

• Benchmark representative generated Rust XML decoders end to end, including tokenization, tag lookup, and field decoding, across small, medium, and large schemas. Compare default match dispatch with --feature phf using the same documents and tag distributions.
• Record throughput in MB/s and elements/s, plus memory use and the hardware counters where available. Explain any difference from the lookup-only results.
• Use [Benchmark] Large-schema compile-time & binary-size comparison (600/1500 elements) deferred from #49 #55's large-schema compile-time and size results when deciding whether an automatic threshold is worthwhile.
• Decide whether phf remains explicit opt-in or becomes an automatic default above a measured threshold. If the latter, implement the selection, an override, documentation, and regression tests.

Related: #49, #55, PR #56.