Cache proof receipts across the proving pipeline

[biryukovmaxim]

Wires the proof-receipt store into the transaction, batch, and aggregate provers so a replay — including a flip reorg back onto the same fork — reuses a cached receipt instead of re-proving. Built on top of proving-pipeline-agg; this PR is just that delta.

Why

The proof-receipt store (keyed by semantic proving coordinates) already exists, but nothing read or wrote it. On a reorg the provers re-collected SMT proofs / lane proofs and re-proved work they had already done. This connects the cache to the pipeline so identical coordinates resolve to the stored receipt.

What

• Processor gains tx_image_id / batch_image_id / aggregator_image_id accessors and an AggregatorArtifact associated type. The tx/batch/aggregator artifacts are now bounded Clone + Borsh{Serialize,Deserialize} so the scheduler can serialize them into the store and hand a cached value back out.
• Receipt I/O on the scheduler handles, served off the proving worker's runtime through the storage read/write workers (never a synchronous store read on the async runtime):
  • ScheduledTransaction ↔ per-tx receipt
  • ScheduledBatch ↔ per-batch receipt
  • ScheduledBundle ↔ aggregate (settlement) receipt — keyed off the bundle's own start coordinate, with its first batch as the storage gateway (the aggregate prover holds no store of its own). The methods are generic over S/P, so the published ScheduledBundle<A> handle stays free of storage generics and the settler queue type is untouched.
• Keying is fork-aware: every key starts with the checkpoint-index Prefix, followed by block_hash, image_id, and (per kind) merge_idx / seq_commit. The same checkpoint on two competing chains yields distinct keys, yet both share the prefix.
• Pruning: the pruning worker drops every cached receipt at a pruned checkpoint in one prefix scan, across all forks and programs (a pruned batch can never be rolled back to).
• Durability ordering: on a cache miss each worker proves, waits for the receipt to be durable, then publishes the artifact, so a crash never leaves a consumed-but-uncached receipt. The transaction worker proves inline (one proof on the GPU at a time) rather than spawning per-tx tasks.
• Metadata: ChainBlockMetadata supplies the block hash via BatchMetadata::block_hash; a test-utils-gated BatchMetadata for u64 stands in for tests. TransactionEffects derives Clone + Default + Borsh.

Tests

New scheduling/scheduler/tests/proof_receipt_cache.rs round-trips the tx/batch/aggregate receipt paths through the storage workers (miss → store → hit), covering the cache-hit path the dev-mode sims don't reach.

./cargo-check-all.sh (clippy -D warnings over the workspace + excluded crates) is clean. Real-proof verification on the GPU box: the cuda test-suite e2e and the sim reorg tests (l2_flow_*_reorgs_*, which exercise the cache-hit replay path) run with --features cuda and RISC0_DEV_MODE off.