fix: report plaintext size/etag for multipart objects in listings

[ServerSideHannes]

Problem

ListObjects resolves an object's plaintext size from the backend object's plaintext-size user metadata — but that metadata is only stamped on the simple PUT path. Multipart objects can't carry it (user metadata is fixed at CreateMultipartUpload, before the final size is known); their plaintext size lives in the .s3proxy-internal/<key>.meta sidecar, which _process_list_objects never read. Listings therefore reported the ciphertext/framed size for every multipart object.

Impact, observed on the ScyllaDB backup cluster: scylla-manager's rclone compares listed sizes to decide what to skip. Every multipart object (all large sstables) always looked changed, so every retry pass and every subsequent backup run re-uploaded the entire multipart dataset. Tables show ~180–199% progress with 0 failed bytes; the daily backup ships ~2× its data, runs 30h+, and dedup only works for small (simple-PUT) files. HEAD already resolves the sidecar correctly — only LIST was broken.

Fix

• _process_list_objects: when head metadata lacks plaintext-size, load the multipart sidecar and report total_plaintext_size plus the same synthetic etag that HEAD and CompleteMultipartUpload return. Raw-size fallback remains for objects with neither (legacy/unencrypted, where the backend size is already correct).
• New process-local LRU (state/attr_cache.py) keyed by (bucket, key, backend etag) caches resolved attributes, so repeated listings (a backup client walking the same directories each pass) skip the per-object round-trips entirely — including the pre-existing per-object HEAD. Overwrites change the backend etag, which invalidates the entry without coordination.
• CompleteMultipartUpload primes the cache from its response etag, so freshly uploaded objects list with no extra reads.

Tests

tests/unit/test_list_multipart_plaintext_size.py:

• sidecar resolution for multipart objects (size + synthetic etag)
• raw-size fallback when no sidecar exists
• simple-PUT user metadata still wins without a sidecar lookup
• repeat listings served from cache (no extra HEAD/GET)
• cache scoped to backend etag (overwrite re-resolves)
• failed HEAD falls back and does not poison the cache
• LRU eviction + empty-etag guard
• end-to-end complete → cache primed (reuses the complete-recovery fixtures)

Expected effect on the Scylla backups: retry passes and subsequent runs dedup instead of re-uploading — roughly halving transfer and bringing the daily run back inside its window.