Add cross-backend parity tests for pandas, Arrow, DuckDB, native, and fallback execution

[parthdongre]

Summary

Add a focused interoperability/parity suite that proves equivalent public semantics across supported source types and execution backends.

Why

FrameVitals 0.2.0 is explicitly source-aware and can route work through pandas, Arrow, DuckDB, native Rust, and pure-Python fallback paths. That flexibility is a major strength, but it also creates a correctness surface that should be guarded directly.

Scope

• Build deterministic fixtures that can be represented as pandas, file-backed Arrow/Parquet, PyArrow, and DuckDB inputs.
• Run representative public APIs across native and fallback execution where applicable.
• Compare exact fields exactly and approximate/sketch-derived fields within documented tolerances.
• Verify provenance accurately describes the execution path.
• Keep the suite small enough for normal CI.

Acceptance criteria

• Representative analyze, profile, compare, validate, and gate paths are covered where source support applies.
• Exact metrics match across backends.
• Approximate metrics use explicit documented tolerances.
• Native-disabled fallback behavior is tested.
• Execution provenance is asserted.
• Failures identify the source/backend pair clearly.

Priority: P1 — correctness / interoperability
Suggested labels: testing, interoperability, arrow, rust