[python] Parse BlobDescriptor v1/v2 bytes without misclassifying inline payload

[Stephen0421]

Purpose

Align PyPaimon's BlobDescriptor wire format with Java, without changing the existing production read loop.

• serialize() always writes CURRENT_VERSION (v2 + magic), matching Java BlobDescriptor.serialize().
• deserialize() accepts v1/v2 prefixes, trailing bytes, and signed URI length (<i, reject negatives).
• parse_if_serialized is an exact-length parse for v1/v2 (used by writer validation).
• from_descriptor_bytes is the explicit API for bytes known to be a descriptor. It uses deserialize() (Java prefix contract, trailing padding allowed). It is not a detector: garbage that happens to look like a v1 prefix can yield a BlobRef with a nonsense URI.
• from_bytes() / is_blob_descriptor stay v2-magic-only so inline blob payload is not classified as a v1 descriptor.
• blob-descriptor-field writes (write() and write_row()) accept exact v1/v2 bytes and reject trailing padding. Version is checked from data[0] before parse (in [1, CURRENT_VERSION]). Writer validation uses exact wire length instead of a serialize() round-trip, because that round-trip would reject legal v1 after serialize() became always-v2.
• blob.stored-descriptor-fields is still not a layout switch. Python master ignored that key and wrote dedicated .blob files; a global fallback would mis-parse those tables on a rolling upgrade. The cost is that Java tables which only set the fallback key store inline descriptors, and Python returns those bytes instead of fetching payload. Use blob-descriptor-field (column directives already copy the legacy key onto the canonical option).

Out of scope (follow-up): production convert readers still call Blob.from_bytes(), so historical v1 descriptor columns are stored successfully but not resolved to payload. That path should use from_descriptor_bytes.

Tests

• pypaimon.tests.blob_test.BlobTest — descriptor serialize/parse, writer v1/v2/trailing/version, from_bytes(allow_blob_data=False) error wrapping
• pypaimon.tests.blob_table_test.DedicatedFormatWriterTest.test_legacy_stored_descriptor_fields_keeps_dedicated_blob_layout
• pypaimon.tests.column_directive_utils_test

[JingsongLi]

[P2] Keep exact VideoFrameDescriptor inputs valid

BlobDescriptor.deserialize() dispatches through BlobDescriptorSerde, so an exact serialized VideoFrameDescriptor is a supported descriptor and the previous round-trip validation accepted it. The new parse_if_serialized() only implements the ordinary v1/v2 BlobDescriptor layouts; for VideoFrameDescriptor("file:///v.mp4", 0, 10, 2).serialize(), deserialization succeeds and reserialization is byte-for-byte equal, but this call returns None, so every such value is now rejected as having trailing bytes. Please dispatch the exact-length check through the serde (or special-case VideoFrameDescriptor with its exact-length deserializer) before applying the ordinary v1/v2 calculation.

[Stephen0421]

Thranks. parse_if_serialized() now dispatches through BlobDescriptorSerde: an exact VideoFrameDescriptor is parsed with its exact-length deserializer first, then the ordinary v1/v2 length check. Writer validation still calls BlobDescriptor.parse_if_serialized(), so VideoFrameDescriptor("file:///v.mp4", 0, 10, 2).serialize() is accepted and padded input is still rejected.