fix(format): reconcile inline blob descriptor BINARY -> LARGE_BINARY on parquet read

[duanyyyyyyy]

Purpose

Reading a column declared with blob-descriptor-field fails with:

Invalid: src type binary and target type large_binary mismatch

Inline blob descriptors are stored in the main parquet file as BINARY, while a
BLOB column's read type is LARGE_BINARY (a BLOB field is large_binary in the
table schema; the writer narrows it to binary for inline storage). In
ParquetFileBatchReader::NextBatch, the per-batch read-schema reconciliation
required the parquet array type to equal the read type, so it rejected this legal
binary/large_binary difference before any blob handling ran — including before
BlobViewResolvingBatchReader, which itself expects a LargeBinaryArray input.

That reconciliation already walks the whole read schema every batch (for timestamp
timezones), so this PR folds the blob widening into it rather than adding a second
pass. ParquetTimestampConverter is renamed to ParquetTimestampBinaryConverter
and, in the same traversal:

• NeedCastArrayForTimestamp treats BINARY -> LARGE_BINARY as "needs cast"
  instead of a mismatch;
• CastArrayForTimestamp widens BINARY arrays to LARGE_BINARY (rebuild the
  32-bit offsets as int64, reuse the value/null buffers — no value-data copy).

Every other column (including timestamps awaiting timezone adjustment) is
unchanged, and non-blob reads do no extra work. This is the read-side inverse of
the large_binary -> binary narrowing the writer already applies to inline blob
fields.

Tests

• parquet_file_batch_reader_test.cpp: adds TestReadBinaryColumnWithLargeBinaryReadSchema
  — writes a {int32, binary} parquet, reads it through a {int32, large_binary}
  schema, and asserts the blob column is widened to large_binary with correct
  values (including null / empty) while the int32 column is untouched.
• Existing timestamp converter tests (renamed to
  parquet_timestamp_binary_converter_test.cpp) continue to pass.

API and Format

No file-format change. Internal only: ParquetTimestampConverter is renamed to
ParquetTimestampBinaryConverter; no public API is affected.

Documentation

No documentation change needed.

Generative AI tooling

Claude Code 4.8

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

duanyan.duan seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}