Support Iceberg v3 nanosecond timestamps in Arrow, manifests and pruning

[osipovartem]

Summary

• Preserve timestamp_ns and timestamptz_ns across Iceberg values, Arrow schema inference, Parquet bounds, manifests, partition transforms, and DataFusion statistics/pruning.
• Use Iceberg-compatible floor-microsecond bucket hashing and widened distance arithmetic for valid extreme nanosecond ranges.
• Keep default SQL CREATE on v2 microsecond timestamps; normalize DataFusion format.format-version and use the parsed format version for explicit v3, including accepted 03.
• Preserve the low-cardinality bitmap path and use indexed, lazy output groups for high-cardinality partitions without rejecting batches above 10,000 partitions.

Verification

• cargo +1.95.0 test -p iceberg-rust-spec -p iceberg-rust -p datafusion_iceberg --lib --quiet: 523 active tests passed.
• cargo +1.95.0 test -p datafusion_iceberg --test roundtrip_types --test integration_df_types --quiet: 11 passed. Includes v3 nanosecond write/read/filter and default-v2 versus explicit-v3 SQL CREATE metadata checks.
• cargo +1.95.0 clippy -p iceberg-rust-spec -p iceberg-rust -p datafusion_iceberg --all-targets --all-features -- -D warnings: passed.
• Focused partition microbenchmark (debug, 10 x 65,536 rows): 4 partitions 148.7 ms base vs 147.2 ms here; 2,048 partitions 284.6 ms here. The temporary benchmark code/worktree were removed.
• Independent read-only review approved after addressing its findings.

Scope

This proves local Iceberg v3 round-trips. Live Snowflake-managed v3 Horizon read interoperability remains to be tested after the Rustice dependency pin is updated; it is not claimed here.