[Bug][Meta Cache] Complete global invalidation ordering after PR #65126

### Background

PR #65126 now fixes the per-key ordering problem between object cache entries and ID-to-name navigation state for external metadata caches.

The current PR covers the key-scoped paths below for both database and table layers:

- hot named lookup
- miss-load / generation-rejected load
- register / update
- invalidate / unregister / rename

So this issue no longer tracks those per-key races.

### Remaining gap

The remaining correctness gap is on full invalidation / reset style paths that still update object caches and ID maps in separate steps, without a full publication barrier across all affected keys / stripes.

Examples include patterns like:

- `tables.invalidateAll(); tableIdToName.clear();`
- `databases.invalidateAll(); dbIdToName.clear();`

With the current structure, a concurrent named lookup / load / publication may still race with `invalidateAll()` / full reset and repopulate object or ID state after cleanup has started.

### Why this is a follow-up issue

Fixing this correctly is broader than the per-key ordering work already merged into PR #65126.

A proper solution likely needs:

- a full invalidation barrier or equivalent global generation protocol
- fixed-order coordination across stripes / keys to avoid deadlock
- careful interaction with manual load, refresh, and removal-listener paths
- dedicated deterministic concurrency tests

That is a larger change than the current PR scope.

### Expected scope

This follow-up should focus on the remaining non-per-key cases, including:

1. full invalidation / reset paths for table caches
2. full invalidation / reset paths for database caches
3. interactions between invalidate-all style cleanup and concurrent named lookup / miss-load / publication

### Expected guarantees

At a minimum, the final fix should ensure:

- `invalidateAll()` / full reset cannot be undone by a concurrent late publication
- object cache and ID navigation state are cleared under a consistent global ordering rule
- the per-key guarantees already added in PR #65126 remain unchanged
- replay stays cache-only
- no remote I/O is introduced under publication locks

### Test expectations

Please add deterministic concurrency tests for at least:

- `invalidateAll()` / full reset vs hot named lookup
- `invalidateAll()` / full reset vs blocked miss-load
- both table and database cache paths

### Related context

- PR #65126
- Review discussion: https://github.com/apache/doris/pull/65126#discussion_r3642519993

Start by tracing the table and database cache paths around invalidateAll(), clear(), named lookup, miss-load, and publication, using PR #65126 and its review discussion for existing ordering rules. Add deterministic concurrency tests for full invalidation versus hot lookup and blocked miss-load on both cache paths. Done means late publication cannot undo cleanup, object and ID state follow one global ordering, replay remains cache-only, and publication locks perform no remote I/O.

Written by the indexing model from the issue text.