perf(gpu): overlap replay with shard proving

[hero78119]

Problem

GPU shard proving serialized compact CPU replay with the preceding shard proof. Replay ownership also crossed shard boundaries implicitly, and the packed producer priority could not represent the full supported producer domain.

Design Rationale

A capacity-one pipeline prepares shard N+1 on the CPU while shard N is proved. CUDA assignment remains on the proving thread and starts only after the preceding proof completes, preventing assignment kernels or transfers from overlapping sumchecks. Replay owners are retained and recycled explicitly so their storage remains valid across the boundary.

Producer base, count, and order remain separate on the host. The hot device domain contains only base and count; order is read only for terminal emission. A checked 64-bit priority preserves continuation class, instruction order, producer index, and slot through exactly 2^28 rows without changing transcript or witness order.

Change Highlights

• ceno_emul: add compact replay ranges and checked producer-domain metadata.
• ceno_zkvm: overlap CPU replay with the preceding proof, retain/recycle owners explicitly, and gate each GPU assignment before proving its shard.
• GPU dispatch: publish initialized shard metadata across streams before dependent kernels run.

Benchmark / Performance Impact

Operation

Operation	master (s)	this PR (s)	Improve (master -> this PR)
Block 23817600 total proof creation	52.095	48.604	3.491 s (6.70%)

Layer

Layer	master (s)	this PR (s)	Improve (master -> this PR)
Application proving	46.144	42.752	3.392 s (7.35%)
Recursion proving	5.882	5.786	0.096 s (1.64%)

Benchmark command:

target/release/ceno-reth-benchmark-bin --mode prove-stark --block-number 23817600 --output-dir output --cache-dir rpc-cache --chain-id 1

Environment: NVIDIA RTX 4090 (CUDA_ARCH=89), GPU + jemalloc + AOT benchmark workflow. Baseline used Ceno 436a32e2; this PR used 6fddf6cb.

Raw data:

• master: run 32851091196
• this PR: run 32947621524

Testing

cargo fmt --all -- --check
cargo check -p ceno_zkvm --all-targets
RUST_MIN_STACK=33554432 cargo test --workspace --lib --bins --tests --examples --no-default-features --features goldilocks --no-run
RUSTFLAGS=-Dwarnings cargo check --workspace --all-targets

Both benchmark workflows completed successfully and verified block 23817600.

Risks and Rollout

The main risk is cross-shard lifetime or stream-ordering regression. Bounded channels, explicit owner recycling, checked domain bounds, and a publication barrier make these boundaries explicit. Reverting the PR restores serialized replay without data migration or protocol changes.

Copilot Reviewer Directive (keep this section)

When Copilot reviews this PR, apply .github/copilot-instructions.md strictly.