fix(gzip): bound decompression memory

[CyMule]

Summary

• Decompress gzip uploads incrementally in 1 MiB chunks.
• Keep decompressed outputs up to 1 MiB in memory and spool larger outputs to disk.
• Proxy disk-backed spools so downstream MIME detection and partitioning do not copy them back into BytesIO.
• Reuse the request-owned UploadFile so FastAPI closes the decompressed temporary file after the response.
• Add coverage for memory-backed and disk-backed outputs, stream cleanup, and proxy file behavior.

Why

ungz_file currently reads the complete expanded payload into a bytes object before wrapping it in BytesIO. Peak Python memory therefore grows with the uncompressed file size.

A raw SpooledTemporaryFile is not sufficient by itself because existing downstream code recognizes that concrete type and copies its complete contents back into memory. Incremental decompression plus a normal file proxy avoids both document-sized allocations.

This does not introduce a file-size limit or change the partition API contract.

Impact

Local end-to-end benchmarks with valid documents produced byte-identical responses:

• A 54.9 MB DOCX reduced median peak process RSS from 828.6 MiB to 759.9 MiB (-68.7 MiB, 8.3%) while median latency changed from 1.348s to 1.370s (+1.6%).
• A 70.4 MB PDF reduced median peak process RSS from 1,029.5 MiB to 951.6 MiB (-77.9 MiB, 7.6%) while median latency changed from 0.532s to 0.529s (-0.6%).

In decompression-and-reread benchmarks, request-attributable RSS fell from 108.1 MiB to 5.3 MiB for the DOCX and from 133.6 MiB to 4.8 MiB for the PDF. That isolated stage was 4.0% and 8.3% slower respectively, but decompression was a small portion of the complete request.

Large expanded outputs use temporary disk space rather than equivalent heap memory.

Risk

Disk-backed outputs add temporary-file I/O. The 1 MiB copy buffer keeps this overhead small while maintaining bounded heap use.

Temporary-disk capacity remains the external bound; this change does not impose a maximum expanded size.

Validation

• uv run pytest test_general/api/test_gzip.py -q — 3 passed, 28 xfailed.
• uv run ruff check --select I prepline_general/api/general.py test_general/api/test_gzip.py — passed.
• Valid 54.9 MB DOCX and 70.4 MB PDF responses were byte-identical before and after.
• Paired decompression-and-reread and complete API benchmarks covered memory-backed and disk-backed behavior.