[DRAFT] perf(2892): preload notification objects, cache assignments

[tuannx]

Problem (upstream IQSS#2892, private repro)

Rendering a notifications page resolves each row's objectId with service find() calls (up to 3 sequential finds for role notifications), plus getEffectiveRoleAssignments per row and requestor loads per row. Measured on a 12-notification/9-object fixture: 71 SELECTs for the legacy per-row pattern.

Fix

• InAppNotificationsJsonPrinter.preload(): partitions notification ids by target kind (from the type switch) and loads everything with a fixed set of IN queries - dv objects, files with full to-one closure, versions, dataset display graphs (versions+fields), file metadatas - plus a per-page cache of effective role assignments keyed by (user, object).
• New batch service methods: DvObjectServiceBean.findDvObjectsByIds, DataFileServiceBean.findFilesByIds, DatasetVersionServiceBean.findVersionsByIds, DatasetServiceBean.preloadDisplayGraphs, DataFileServiceBean.preloadFileMetadatas. To-one fetch closures only: no cartesian products. Inverse-side collections are set, never join-fetched (verified zero writes).
• JsonPrinter.json loop preloads once per page; new 4-arg addFieldsByType resolves from the preload with identical JSON branches (old 3-arg delegates for single use).
• UserNotificationServiceBean.findByUser fetch-joins requestors (11 -> 7 on the fixture; delta is exactly the 4 requestor loads).

Verification

• New NotificationRenderBudgetIT (real services wired with the test EM, incl. the permission/group chain): legacy-vs-preloaded 71 -> 50, repeated notifications add ~zero queries (51 vs 50, 1 query of documented L1-order tolerance), JSON content asserts per type, zero writes.
• No new checkstyle violations in touched files.

Notes / follow-ups

• Residual per-object costs (EAGER hydration of thumbnail/quota/theme back-references, per-distinct-object assignment lookups, per-render assignee lookups) are pre-existing and unchanged; the win scales with notification count (repeats are free).
• Measured finding: @BatchFetch does NOT reroute EAGER-at-build loading with weaving disabled (24 queries before and after annotating), so to-one EAGER residuals need fetch joins, not batch annotations. Relevant when reviewing upstream Faster file and dataset pages for datasets with many files IQSS/dataverse#12740's to-one annotations.
• True bulk assignment resolution (one query per page instead of per distinct object) would need permission-system surgery and is left as a follow-up.
• Stays on the private fork until reviewed; upstream proposal to follow.