perf(decode): evaluate lexical crate for float parsing

[membphis]

Background

src/decode/number.rs uses <f64>::from_str (via the standard library) to parse floating-point numbers. The standard library implementation is correct but not optimized for throughput; the lexical crate offers a significantly faster parser with the same correctness guarantees.

Proposal

Profile from_str against lexical::parse::<f64> on a representative mixed-payload bench (many numeric fields). If from_str is measurable in the profile, replace it with lexical-core (no-std compatible, no allocation).

Acceptance criteria

• Benchmark shows lexical is faster in the target workload (many float fields).
• cargo test --no-default-features still passes (lexical is scalar-safe).
• No change to the external API or error semantics — QJD_PARSE_ERROR still returned on malformed numbers.

Notes

lexical-core is a popular, well-maintained crate used by serde_json and simd-json. Adding it is low-risk. If profiling shows from_str is not on the hot path for typical "parse + read 3 fields" workloads, close this as not worth the dependency.