Periodic incremental image_scales refresh: rewrite metadata for objects changed since the last run

[jensens]

Problem

image_scales catalog metadata freezes whatever URL was resolvable at index time. Several routine situations leave entries that are stale or suboptimal, and nothing brings them back on its own:

1. Fresh uploads can never bake a direct Thumbor URL. reindexObject() runs inside the creating transaction, where the image blob has no committed TID yet — ZODB assigns OID/serial during commit. get_blob_ids() therefore returns None (it explicitly rejects z64 serials), _build_thumbor_url() returns None, and the uid fallback URL gets indexed instead.
2. Out-of-request indexing produces @@images URLs rather than Thumbor URLs — see image_scales metadata gets @@images URLs (not thumbor) when indexed outside a request #14.
3. Blob replacement changes a blob's TID. Any already-indexed direct Thumbor URL for that object points at the previous TID.

Since #17, uid URLs heal on traversal, so none of this is broken — but every healed request costs a 302 through Plone, which is precisely the hop image_scales exists to avoid.

What is missing

A maintenance job that periodically refreshes image_scales for only the objects that changed since the last run, rather than the all-or-nothing full walk that exists today.

Rough shape:

• select candidates by modification watermark (not a full catalog walk)
• reindex image_scales only — no other indexes, no object rewrite
• chunked and resumable, with the watermark persisted per run
• safe to run unattended on a site with ~139k objects

Relationship to other issues

• image_scales metadata gets @@images URLs (not thumbor) when indexed outside a request #14 — root cause for the out-of-request variant. Fixing it removes one source of stale entries but not the fresh-upload one, which is inherent to the commit ordering.
• purge_scales: unusable on large sites (reindexes every object, times out, no chunking, missing pgthumbor layer in zconsole path) #16 — purge_scales needs chunking and resumability for the same reason; the traversal/batching machinery should be shared rather than written twice.
• Upcoming Thumbor source derivative work (downscaled, colour-normalised source blob chosen in preference to the original): derivatives likewise only receive a TID after commit, so a backfill has to reindex in a later pass. That spec deliberately scopes the periodic job out, on the assumption it is solved here.

Out of scope

Changing when or how URLs are generated. This is purely about keeping already-correct generation reflected in the catalog over time.