Add opt-in ordered Snowflake CSV schema inference

[osipovartem]

Summary

• Add an opt-in Snowflake CSV inference mode. Default Arrow CSV inference and the existing order-independent exact-decimal mode are unchanged.
• Preserve the first observed type and compact type/precision state so callers can merge streamed chunks in order without reparsing rows.
• Match observed asymmetric transitions: NUMBER then REAL becomes TEXT, REAL then NUMBER stays REAL; DATE then TIMESTAMP stays DATE, the reverse becomes TEXT.
• Clip combined fixed-point scale to fit precision 38 while leaving individually overflowing numeric values in the REAL family.

Evidence

• Narrow Snowflake 10.36.101 probes checked both row orders for scientific notation, 39-digit integers, scale-38 fractions, and DATE/TIMESTAMP. Separate probes confirmed 37-digit integer + scale 2 -> NUMBER(38,1) and 38-digit integer + scale 1 -> NUMBER(38,0).
• cargo +1.95.0 test -p arrow-csv --lib: 99/99 passed.
• The ordinary path uses compile-time dispatch; no extra per-row classifier or materialization is added to default CSV reads.
• Independent read-only review: approved; the reviewer also checked 1,000 three-chunk sequences under both merge groupings.