fix(transform): apply truncate to binary values

[jaideeppyne]

Which issue does this PR close?

• Closes truncate transform rejects binary values in transform_literal and LargeBinary arrays #3119.

What changes are included in this PR?

The spec lists binary under the source types for truncate[W], and the Truncate Transform Details table gives v.subarray(0, L). Transform::result_type already accepts PrimitiveType::Binary, so a binary truncate partition field is legal, but neither half of TransformFunction for Truncate could actually handle one:

• transform_literal had no PrimitiveLiteral::Binary arm, so Transform::project and strict_project returned FeatureUnsupported for every operator. InclusiveProjection in scan/cache.rs propagates that, so a scan with a filter on such a column errors instead of pruning.
• transform handled DataType::Binary but not DataType::LargeBinary, and schema_to_arrow_schema maps Iceberg binary to LargeBinary. So the array path failed on the exact arrow type this crate produces for a binary column.

truncate_binary already existed and was correct, it just was not reachable from either entry point. I added the two arms. fixed stays rejected, since the spec does not list it as a truncate source type, and the literal arm is guarded on the datum type so a Fixed datum (also backed by PrimitiveLiteral::Binary) still errors.

Are these changes tested?

Yes, unit tests in truncate.rs: array truncation for BinaryArray and LargeBinaryArray, literal truncation (longer than width, shorter than width, empty), inclusive projection for </<=/>/>=/=/STARTS WITH/IN, strict projection for !=, and that fixed is still rejected. Reverting just the two source arms and keeping the tests fails the 4 new tests and leaves the 16 pre-existing truncate tests passing. Full cargo test -p iceberg is green (1617 tests), fmt and clippy clean.

I found this with a differential harness over bucket[N], truncate[W], year/month/day/hour on a 781-value corpus (spec test vectors plus randoms and boundary values across every source type), running the same inputs through pyiceberg 0.11.1, iceberg-go and this crate's literal and array paths. Of 9333 comparisons, 49 were binary truncate literals where pyiceberg and iceberg-go return a value and this crate errored. The LargeBinary half is not visible that way, I hit it after checking what schema_to_arrow_schema actually emits for binary.

Not covered: BinaryView, which no path in this crate currently produces, and which bucket does not handle either.

AI Disclosure

Claude Code wrote the differential harness and drafted the patch and tests. I checked the spec clauses, verified every claim above against real test output, and reviewed the final diff.