[Testing] Parquet f32 tests and benchmark

[0lai0]

What

Add fidelity tests and an optional benchmark script that verify correctness and measure throughput for the end-to-end f32 Parquet pipeline enabled by A1–A3.

Why

Without dedicated tests, regressions in the f32 reader or dispatch path may go undetected. Without a benchmark, the expected ~25–35% throughput improvement over the f64 path (when the Parquet column is native f32) remains unvalidated.

How

• Fidelity test: compare f64-kernel output vs f32-reader + f32-kernel output (threshold ≥ 0.99999 where applicable) in qdp-core/tests/
• Optional: qdp-python/benchmark/benchmark_parquet_f32.py reporting vec/s for f32 vs f64, with a usage comment or README note
• Tests run in CI Linux CUDA job (or CPU smoke if no GPU available in CI)

Acceptance criteria:

• Tests run in CI Linux CUDA job (or CPU smoke if no GPU in CI)
• Benchmark script has usage comment or README note