Add recoverable CSV record error handler

[osipovartem]

Summary

• add an opt-in CsvRecordErrorHandler to the Arrow CSV reader
• report line, byte offset, expected/actual field counts, and original record bytes
• skip malformed field-count records only when a handler is configured
• preserve the existing allocation-free strict decoder path when no handler is configured
• bound recovery-only field-offset growth

Why

DataFusion consumers implementing Snowflake-compatible COPY ... VALIDATION_MODE need to report malformed CSV rows and continue scanning later valid rows. Arrow currently aborts on the first field-count mismatch and discards the source record metadata.

Performance

The default path dispatches once per input chunk and then executes the existing strict decoder loop unchanged. Raw record retention and recovery allocations are enabled only when a handler is explicitly configured.

Tests

cargo +1.95.0 test -p arrow-csv --lib --offline

73 tests passed, including chunk-boundary recovery for both extra and missing fields.

[osipovartem]

Performance evidence: Criterion on the existing 4-column, 4,096-row CSV reader benchmark measured 329.91 us before and 333.16 us after (+0.96%). Criterion classified the result as within the noise threshold. The no-handler path retains the original strict decoder loop; recovery allocations remain opt-in.