fix(gzip): stop holding decompressed uploads in memory

[CyMule]

What

Gzip uploads are decompressed in 1 MiB chunks into a temporary file that stays in memory up to 1 MiB and spills to disk above that. Previously the whole expanded payload was held in memory.

Before: .gz upload → gzip.read() → bytes (full size, in RAM) → new UploadFile (never closed by the request)
After:  .gz upload → 1 MiB chunks → SpooledTemporaryFile (disk above 1 MiB) → same request-owned UploadFile

• The decompressed file sits on the request's own UploadFile, so FastAPI closes it when the request finishes. Partial output is closed if the gzip data is corrupt or truncated.
• No size limit is introduced, and the partition API is unchanged.

Why

ungz_file held the expanded payload as one bytes object for the whole request, so peak memory grew with the uncompressed size.

An earlier revision passed large outputs downstream behind a proxy object, to stop unstructured from copying them back into memory. That proxy made convert_to_bytes() reject the file: ValueError: Invalid file-like object type for gzipped uploads over 1 MiB on the PDF OCR and text-encoding-detection paths. This revision passes the spool itself. Avoiding the downstream copies now comes from unstructured#4419 (open; expected to ship as 0.27.12).

Merge dependency

Stacked on #579. Released as 0.1.13, after #579's 0.1.12. The decompression-stage savings apply on any unstructured version. The end-to-end savings below need the unstructured#4419 release (expected 0.27.12); this PR does not change the lock. On 0.22.18 the DOCX/PPTX/PDF partitioners still copy the spool into memory.

Impact

Measured on the earlier revision, whose downstream path matches unstructured#4419. Responses were byte-identical:

• 54.9 MB DOCX: peak RSS 828.6 → 759.9 MiB; latency 1.348s → 1.370s.
• 70.4 MB PDF: peak RSS 1,029.5 → 951.6 MiB; latency 0.532s → 0.529s.

Risk

A very large expansion now uses temporary disk instead of RAM. Disk capacity is the external bound, and a decompressed-size cap would be a product decision for the upload-limits work.

Validation

• test_gzip.py + test_filetypes.py pass on unstructured 0.22.18, including an end-to-end gzipped request that asserts the decompressed file is closed after the response (it fails against main's ungz_file). The same tests pass with the unstructured#4419 branch installed over the locked dependencies. That includes a regression test that reads a disk-backed decompressed upload through convert_to_bytes(); it fails with the error above on the earlier proxy revision.
• Ruff passed.

[cubic-dev-ai]

All reported issues were addressed across 3 files

<sub>Shadow auto-approve: would not auto-approve because issues were found.

Re-trigger cubic</sub>

[cubic-dev-ai]

0 issues found across 3 files (changes from recent commits).

<sub>Shadow auto-approve: would not auto-approve. Auto-approval blocked by 1 unresolved issue from previous reviews.

Re-trigger cubic</sub>

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[cubic-dev-ai]

No issues found across 4 files

<sub>Shadow auto-approve: would not auto-approve. This PR does not meet the repository auto-approval settings.

Re-trigger cubic</sub>