[Refactor] Add FloatElem trait and DataReader<T>

[0lai0]

What

Generalize DataReader from Vec<f64> to Vec<T> by introducing a FloatElem trait, so that dtype=f32 data from file sources can stay as f32 through the entire read path.

Why

DataReader::read_batch returns Vec<f64> only (qdp-core/src/reader.rs; the comment at line 100 marks this as structural debt). Until this is fixed, dtype=f32 file sources are silently cast to f64, preventing a true f32 pipeline.

How

• Add FloatElem marker trait with f32 / f64 impls (or reuse existing Dtype mapping) in qdp-core/src/types.rs or a new qdp-core/src/float_elem.rs
• Change DataReader → DataReader<T: FloatElem>, read_batch -> Result<(Vec<T>, usize, usize)>
• Update StreamingDataReader<T> the same way
• Fix call sites that no longer compile (minimal — non-Parquet readers may still produce f64 for now)
• No behavior change on existing f64 paths; trait docs explain the f32 end-to-end goal

Out of scope: Parquet schema/cast logic, pipeline_runner wiring
Acceptance criteria:

• cargo test -p qdp-core passes on Linux
• Trait docs explain f32 end-to-end goal
• No behavior change on existing f64 paths