[C++][Parquet] Fix endianness and test failures on s390x (big-endian)

[k8ika0s]

Describe the bug, including details regarding any error messages, version, and platform.

Describe the bug

Running Apache Arrow / Parquet 22.0.0 on s390x (big-endian) exposed multiple endianness-related correctness bugs and some allocator/test fragility:

• Parquet encoders sometimes emit host-order bytes instead of canonical little-endian.
• Decoders may double-swap values or interpret host-order statistics incorrectly.
• INT96 timestamp handling mixes host-order and LE limbs, breaking comparison and decoding.
• Page index min/max expectations are built from host-order bytes.
• Float16 (half) serialization does not always produce canonical bytes.
• BYTE_ARRAY and dictionary 4-byte length prefixes are encoded inconsistently across paths.
• Bit-packing, VLQ, and ZigZag helpers assume little-endian layout and misbehave on big-endian.
• Synthetic OOM tests can trigger allocator-level aborts on large-memory s390x systems (e.g. mimalloc) before Arrow can raise OutOfMemory.

Together these can lead to:

• Non-portable or corrupted Parquet pages and statistics on big-endian hosts.
• Incorrect page index min/max and statistics (wrong sort order, incorrect filtering).
• Mis-decoded INT96 timestamps.
• Crashes in test OOM paths on s390x.

To Reproduce

On an s390x (big-endian) system:

1. Build Arrow C++ (22.0.0) with Parquet enabled.
2. Run relevant gtest suites (e.g. parquet-reader/writer, parquet-arrow reader/writer, page index tests, chunker/Float16 tests, bit-utility tests, IPC message tests, and allocator/buffer OOM tests).
3. Observe:
   • Parquet tests failing due to mismatched bytes (stats, page index, INT96, Float16, RLE prefixes, dictionary pages).
   • IPC byte-identical test comparing against a macOS/ARM snapshot that does not match s390x output.
   • OOM tests aborting when mimalloc is used on large-memory s390x environments.

Expected behavior

• Parquet encoders/decoders should consistently read/write canonical little-endian bytes for all primitives and metadata, independently of host endianness.
• INT96 timestamps, page index min/max, Float16, and binary/dictionary pages should round-trip correctly on s390x and match the expected canonical layout.
• IPC “byte identical” tests should compare against a canonical snapshot that matches big-endian builds as well.
• Synthetic OOM tests should exercise Arrow’s OutOfMemory behavior without provoking allocator-level hard aborts.

What this PR changes

The associated PR:

• Introduces an ARROW_ENSURE_S390X_ENDIANNESS CMake option (default ON on s390x) so Parquet I/O always uses explicit little-endian conversions on big-endian hosts.
• Adds parquet/endian_internal.h with portable helpers for encoding/decoding primitives (integers, floats, INT96, etc.) in little-endian format regardless of host architecture.
• Updates Parquet encoders/decoders (Plain, ByteStreamSplit, Dictionary, DeltaBitPack, FLBA) to use these helpers, including Arrow array fast-paths.
• Ensures level headers, RLE prefixes, and BYTE_ARRAY length prefixes are always written/read as little-endian.
• Canonicalizes INT96 timestamp handling and test fixtures to little-endian limb layout.
• Fixes page index and Float16 expectations so tests compare against canonical LE bytes.
• Adjusts geospatial WKB parsing and related tests to respect a unified endianness shim.
• Updates IPC byte-identical tests to use a canonical s390x snapshot and emit hex diffs on mismatch.
• Clamps synthetic OOM allocations and skips mimalloc-specific OOM tests to avoid allocator aborts while still exercising Arrow’s OutOfMemory behavior.

Environment

• Architecture: s390x (big-endian)
• Component(s): C++, Parquet
• Arrow version: 22.0.0 (and current main while developing the fix)
• Allocators: default pool, mimalloc (where OOM tests were unstable)

Additional context

The goal is to make Arrow/Parquet fully correct and stable on s390x by enforcing canonical little-endian Parquet I/O and adjusting tests so that all suites pass on big-endian platforms without behavior differences or crashes.

Component(s)

C++, Parquet

[kou]

@Vishwanatha-HD created #48151 and sub-issues and opened PRS for them. Is this duplicate of them?

[k8ika0s]

I don't know @Vishwanathan-HD. This is my own work, done without any knowledge someone else was about to put up a PR for what I had been working on.

Also for context, here is the detailed technical write-up of the s390x issue discovery, testing methodology, and architecture-level design approach:

<Apache Arrow & Parquet on IBM s390x>

This PR is the implementation of that research.

Even if some parts end up merged separately or differently, I'm happy to collaborate and ensure the final s390x solution is consistent, canonical, and maintainable across the Arrow/Parquet C++ codebase.

[kou]

OK. Could you review PRs from @Vishwanatha-HD instead of #48212 ?
#48212 has many unrelated changes. We can't review #48212.

[k8ika0s]

Thanks, @kou — that makes sense.

The big diff in #48212 is mostly because I pulled in fresh upstream commits before adding my endianness work, so it looks a lot noisier than it really is. But I totally get that reviewing a giant PR like that isn’t practical.

I’ll start going through the PRs from @Vishwanatha-HD and add technical feedback there, especially around keeping the behavior consistent for canonical little-endian I/O on s390x.

If you or the other maintainers eventually want me to break out pieces of my work into smaller, focused PRs, I’m happy to do that — just let me know the direction you prefer and I’ll follow it.

Appreciate the guidance!

[kou]

Thanks. If you find missed changes in the PRs from @Vishwanatha-HD, could you open a separated (smaller) issues/PRs for them?

Can we close this in favor of #48151?

[k8ika0s]

@kou Sounds good
Closing this and will continue to provide support in other related PRs/Issues.