Pin Arrow CSV blank-row and source-position integration

[osipovartem]

Summary

• Pass the reviewed Arrow CSV blank-record policies and per-file Skip counters through DataFusion CSV scans.
• Preserve intra-file repartitioning for strict policies when ranges are safe (no handler, physical-header skip, or multiline values). On a later-range decoder error only, replay the same file once to recover file-relative diagnostics; use ETag If-Match when available and keep replay out of Skip counters.
• Reject protobuf serialization if it would silently discard policy or counter state. Keep the default Ignore path unchanged.
• Rustice #661 sets newlines_in_values for quoted COPY input so quoted multiline records remain unsplit.

Verification

• DataFusion CSV library: 28 passed, 1 ignored benchmark; includes LF/CRLF splits, EmitNullRow, per-file Skip counters, and replaced-object replay.
• CSV crate Clippy: passed with no warnings in the changed crate (dependency warnings remain).
• Ignored 64 MB in-memory scan benchmark: 61.8 MiB/s serial vs 420.6 MiB/s over eight ranges (median of three). This is scan-level, not end-to-end COPY throughput.
• Rustice #661 repinned to 02d5697: targeted COPY blank-line fixture 55/55; full bronze corpus 3217/3217 with zero baseline regressions.
• Independent read-only review: approved with no blocking correctness, performance, or API findings.