[Testing][QDP] f32 Parquet fidelity tests and benchmark

[0lai0]

Related Issues

Closes #1342
Part of #1338

Changes

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Why

The f32 Parquet pipeline (A1–A3) lacked dedicated end-to-end fidelity coverage and throughput validation. Without tests, regressions in the f32 reader or dispatch path could go undetected, and without a benchmark, the expected 25–35% throughput improvement over the f64 path for native f32 columns remained unmeasured.

How

• qdp-core/tests/parquet_f32_fidelity.rs: End-to-end f32-vs-f64 fidelity test for the Parquet pipeline.
  • CPU smoke(asserts in CI): f32 and f64 readers return the same logical values up to f64→f32 rounding.
  • GPU fidelity (Linux + CUDA, self-skip on no-CUDA stub via kernel probe): same deterministic data written as native List<Float32> / List<Float64>, read through matching readers, encoded with f32/f64 kernels, compared via fidelity_cross_precision. Thresholds match gpu_fidelity.rs (1e-3 at 8–16 qubits; amplitude f32 norm precision — the issue's "≥0.99999 where applicable" does not apply here).
• qdp-core/tests/common/mod.rs: Shared List<Float32> / List<Float64> Parquet writers for fidelity fixtures.
• qdp-python/benchmark/benchmark_parquet_f32.py + README.md: Optional local/GPU throughput benchmark (vec/s + f32/f64 speedup). Default run generates separate native f32 and f64 files so the f32 path exercises zero-copy native read; --parquet PATH benchmarks a single user file at both dtypes. Not wired into CI.
• .github/workflows/python-testing.yml: Run parquet_f32 (existing [Feature] ParquetReader: f32/f64 columns + Arrow cast #1340 reader acceptance) and parquet_f32_fidelity in the rust-check job under QDP_NO_CUDA=1. CPU smoke asserts; GPU cases skip without a GPU. Scoped to these two binaries — other gpu_*.rs tests lack probe-skip and would fail on the stub.

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[0lai0]

cc @rich7420

[rich7420]

@0lai0 thanks for the patch!
LGTM overall
pulled it down, tests pass and clippy's clean, GPU cases self-skip nicely. Threshold matching gpu_fidelity.rs (1e-3, not 0.99999) is the right call.

Two small nits, neither blocking:

1. write_list_parquet_f32/write_list_parquet_f64 are near-identical (only the builder type differs). Not worth forcing a generic over Arrow builders, but figured I'd flag the dupe.

2. In the benchmark, --parquet vec/s divides metadata num_rows by the time to iterate total_batches batches. Fine for the generated case where they line up, but a bit off for an arbitrary user file where batches*batch_size != num_rows.