feat(serde): Coerce narrow encoding to widened type during PrestoSerializer deserialization

[yingsu00]

When PushDownWidenCast rewrites a consumer fragment's source operator (TableScan or RemoteSourceNode) to declare a wider type, the producer fragment still emits narrow-typed bytes on the wire (its outputLayout stays narrow). The native worker's Exchange operator deserializes those pages via PrestoVectorSerde and must coerce narrow->wide on receive.

Before this commit, the deserialization path threw on the encoding mismatch:
"Serialized encoding is not compatible with requested type: BIGINT.
Expected LONG_ARRAY. Got INT_ARRAY."

This commti patches two call sites of checkTypeEncoding, mirroring the two-pass deserialization in readTopColumns:

1. readColumns (main read pass): Add tryReadWidenedColumn which reads the size + null header, then dispatches to readWideningValues <SourceT, TargetT, Converter> to read narrow bytes from the stream and write widened values into the result FlatVector.

2. readStructNullsColumns (null-tracking pre-pass, only runs when hasNestedStructs(childTypes) is true): Add tryStructNullsSkipWidened which skips numValues * sizeof(SourceT) bytes. Critical: without this, the pre-pass threw before the main pass ever ran on queries whose row schema includes nested ROW types (real-world complex JOIN / aggregate plans hit this constantly).

Supported widening pairs (producer encoding -> consumer kind):

• BYTE_ARRAY (TINYINT) -> SMALLINT, INTEGER, BIGINT
• SHORT_ARRAY (SMALLINT) -> INTEGER, BIGINT
• INT_ARRAY (INTEGER) -> BIGINT
• INT_ARRAY (REAL) -> DOUBLE
• INT_ARRAY (DATE) -> TIMESTAMP (days * kSecondsInDay)

Ordering matters: the widening check is inserted before the existing tryReadNullColumn check in readColumns. tryReadNullColumn consumes bytes from the stream as it tries to interpret the column as UNKNOWN-all-null; if it returns false the stream is past where the widening reader expects to start. Widening goes first; tryReadNullColumn keeps its original spot.

Tests

velox_serializer_test_PrestoSerializerTest adds eight new tests covering each widening pair plus nulls, the nested-struct two-pass path, and a negative test:

• wideningCoercionTinyintToBigint
• wideningCoercionSmallintToInteger
• wideningCoercionIntegerToBigint
• wideningCoercionRealToDouble
• wideningCoercionDateToTimestamp
• wideningCoercionWithNulls
• wideningCoercionWithNestedStructInRow (exercises the pre-pass)
• wideningCoercionUnsupportedPairStillThrows (BIGINT -> INTEGER is
  narrowing, must still throw)
  All pass under all 6 compression-kind test parameterizations (48 total
  test runs).

[xin-zhang2]

should timezone be considered here?

[yingsu00]

Good catch! I have fixed it. DATE→TIMESTAMP converter now does Timestamp::fromMillis(days * 86'400'000) then timestamp.toGMT(*zone) if opts.sessionTimezone is non-null, exactly what CastExpr::castFromDate does.

[xin-zhang2]

Could this introduce a risk that a type encoding mismatch that previously throw an error by checkTypeEncoding is now incorrectly processed silently?
For example, if the serialized data is REAL that is encoded as kIntArray and columnType->kind() is BIGINT.

[xin-zhang2]

@yingsu00 I left a few comments. Please take a look. Thanks.