Fix LZ4_RAW heap decompression failure on chunked BytesInput (#3478)

[yadavay-amzn]

What changes were proposed?

Eagerly materialize the decompressed stream for LZ4_RAW in CodecFactory, matching the existing pattern used for ZSTD. Added a regression test.

Closes #3478

Why are the changes needed?

When reading LZ4_RAW-compressed data through the heap codec path, decompression fails if the decompressed page exceeds ~8KB. The lazy StreamBytesInput.writeInto() path reads via Channels.newChannel() in ~8KB chunks, but LZ4_RAW requires all compressed input in a single buffer for one-shot decompression. This causes MalformedInputException: all input must be consumed in production reads, particularly during dictionary filter evaluation (DictionaryPageReader.reusableCopy()).

The fix extends the existing ZSTD eager-materialization pattern (added for parquet-format#398) to also cover LZ4_RAW:

if (codec instanceof ZstandardCodec || codec instanceof Lz4RawCodec) {
  decompressed = BytesInput.copy(BytesInput.from(is, decompressedSize));
  is.close();
}

How was this tested?

Added testLz4RawHeapDecompressorCanCopyLargePage to TestCompressionCodec:

• Compresses a 16KB random page via CodecFactory heap path
• Decompresses and calls BytesInput.copy() to exercise the chunked materialization path
• Fails without the fix, passes with it

All existing tests continue to pass (6/6 in TestCompressionCodec, 11/11 across all codec tests).