Implement versioned resource providers and revisioned GPU caches

[zoeyrose]

Outcome

Load, validate, upload, cache, and invalidate graphics/font/material resources from packaged, authenticated client, editor-local, and test sources without giving renderer filesystem/network ownership.

Scope and invariants

• Define stable resource IDs, content digests, revisions, type/size/dimension/format limits, dependencies, and missing/error diagnostics.
• Providers supply verified immutable bytes/decoded CPU assets asynchronously; renderer owns bounded decode/upload queues and GPU lifetime.
• Separate CPU cache, GPU cache, transient staging, and consumer manifests with explicit budgets/eviction.
• Track dependencies so one changed editor asset invalidates only affected decoded/GPU/material/scene data.
• Use placeholder/failure presentation that is deterministic and never treats corrupt bytes as a partially valid resource.

Acceptance criteria

• Tests cover missing/corrupt/oversized/wrong-type assets, revision races, provider cancellation, cache eviction, device recreation, hot reload, and duplicate requests.
• Authenticated client cache data is verified before provider publication; editor-local data cannot enter connected-client trust scope.
• Metrics expose hit/miss, bytes, uploads, evictions, revisions, and queue latency with bounded labels.
• Whole-corpus resource manifests fit explicit CPU/GPU/disk budgets.
• No parser for legacy map/content syntax enters renderer crates.

Dependencies and parallelization

Depends on #2/#5. Client resource download and editor project watching can develop separate providers concurrently.

Licensing

Rendering code, shaders, tests, and newly authored fixtures are MIT. Reuse from atrinik/classic sources is allowed only through the audited approved-grantor grant; all other behavior is independently implemented. Bundled assets require exact compatible licenses and notices.