Add opt-in exact decimal CSV schema inference

[osipovartem]

Summary

• Add an opt-in CsvFormat::infer_schema_with_decimal method for one CSV object; the existing FileFormat inference path is unchanged.
• Merge exact Decimal128 precision/scale across streaming chunks, using Arrow's chunk-local zero-integral metadata to avoid spurious Float64 at scale 38.
• Keep mixed numeric/text inference deterministic, and retain a bounded, single-pass stream over the object.
• Pin Arrow to the merged Embucket CSV decimal inference changes.

Validation

• cargo +1.95.0 test -p datafusion-datasource-csv --lib --quiet: 10/10 passed, including multi-chunk decimal, scientific notation, text, zero-only, and 38-digit scale boundaries.
• cargo +1.95.0 clippy -p datafusion-datasource-csv --all-targets --quiet: passed; only a pre-existing warning in datafusion-common remains.
• cargo +1.95.0 fmt -p datafusion-datasource-csv and git diff --check: passed.
• Independent read-only review: approved after addressing two correctness findings.

The opt-in path does not add row-level work to the default CSV inference or scan paths.