[Interleaved IO] Allow filter stages to propagate materialized binary content to downstream stages

Problem

When materialize_on_read=False (the previous default), WDS pipelines with a filter stage suffer a double-fetch: images are read from the tar archive twice.

Stage-level timings on a single 79MB WDS shard (164 images, 56 samples):

Root cause: InterleavedAspectRatioFilterStage decodes each image from the tar to compute width/height, but does not write the decoded bytes back into the binary_content column of the output DataFrame. So image_rows_missing_binary remains 164/164 after the filter, and the writer opens the tar a second time for every image row.

On 80 shards (~10GB), this wastes ~350s of redundant I/O.

Proposed fix

Allow filter stages to persist materialized binary content so downstream stages don't re-fetch. Options:

Option A — Filter writes binary back : After decoding an image to compute aspect ratio, store the decoded bytes in binary_content. Rows that pass the filter already have their binary populated; downstream stages find image_rows_missing_binary=0.

Option A is the simplest: the filter already has the bytes in memory when it decodes for dimension checks — it just needs to populate binary_content before returning the filtered batch.

Impact

• Affects any pipeline: WDS source → filter stage → writer
• No impact on Parquet source pipelines (binary already embedded)

Labels

enhancement, interleaved, performance