fix: prevent skipped model upgrades across task merges and rollbacks

[Rob-Morris]

Problem

• On cold replay and chunk retry, Torii could miss additive model upgrades and leave the indexer database in a poisoned state.
• One failure mode was deserialising historical events against stale schema state, producing PrimitiveError(InvalidEnumSelector) when a payload used a discriminant that only existed in the upgraded schema.
• The other failure mode was advancing model metadata ahead of the backing SQLite schema, leaving models.schema and /schema out of sync with the actual table columns.
• That poisoned state caused follow-on write failures (table ... has no column named ...), read failures (no such column), and retries against stale in-memory state.

Cause

• Historical replay tasks could lose upgrade dependencies when later same-entity work was merged into an existing task, allowing post-upgrade events to run against the old cached schema.
• After rollback, queued SQL such as ALTER TABLE ... could be undone while the in-memory model and token-registration caches remained ahead of committed SQLite state, so retry logic could silently skip required upgrade work.
• Both hazards are present in every torii release since v1.5.0 (2025-04-29).
• They remained latent until cold replay crossed chunks that contained both a schema upgrade and later same-chunk data that required the upgraded schema.

What Changed

• Preserve dependencies when tasks are merged or registered late.
• Use storage-backed model lookup at processor call sites so namespace-filtered missing models remain skippable while real storage failures surface.
• Reset rollback-sensitive model, token-registration, and ERC diff cache state back to committed storage after rollback.
• Add regression coverage for both rollback recovery paths.

Scope

• No fetcher changes.
• No parser-behavior widening.
• No broad cache architecture rewrite.
• Only the two confirmed bug classes and direct regression coverage.

Review-Relevant Side Effects

• Processor model lookup no longer relies on the in-memory model cache alone; on cache miss it now falls back to committed storage and repopulates the cache.
• Processor model lookup no longer treats every storage error as “model missing”; unexpected storage failures now surface instead of being silently skipped.
• Tasks registered late or merged into existing work now retain their prerequisites, which can delay execution until dependencies are satisfied instead of running prematurely.
• Rollback now clears the in-memory ERC balance and total-supply diff before retry, so failed chunks do not carry stale balance deltas forward.
• Rollback now clears cached model state before retry, so model upgrade work may be replayed instead of being skipped under stale in-memory schema state.
• Rollback now rebuilds the token-registration cache from committed storage before retry, so token registration work from a failed chunk may be retried instead of being suppressed by stale in-memory registration marks.

Validation

• cargo test -p torii-task-network
• cargo test -p torii-cache
• cargo test -p torii-sqlite model_optional
• cargo check -p torii-cache -p torii-storage -p torii-sqlite -p torii-processors -p torii-indexer
• cargo test -p torii-cli --lib
• cargo test -p torii-indexer test_rollback_replays_model_upgrade_after_cache_reset -- --nocapture
• KATANA_RUNNER_BIN=/path/to/katana cargo test -p torii-indexer test_rollback_resets_token_registry_for_retry -- --nocapture