Add bounded CSV record filtering to recovery decoder

[osipovartem]

Summary

• Add an opt-in two-stage record filter to the CSV recovery decoder. Valid rows avoid physical-line bookkeeping; the strict/vectorized CSV path is unchanged.
• Bound raw recovery record accumulation when a handler opts in, preventing unbounded buffering for unterminated quoted records.
• Add chunked source-line and record-limit tests plus ignored local throughput benchmarks.

Verification

• cargo test --locked --offline -p arrow-csv --lib: 131 passed, 8 ignored.
• Focused 100k-row quoted CSV benchmark: existing recovery handler ~17-19 ms; filtered handler ~18-19 ms. This is the quote-enabled recovery path only.
• Independent read-only review: approved, no blocking findings.

Dependent DataFusion/Rustice changes will pin this exact commit and add Snowflake COPY diagnostics.