Synchronize concurrent GPU chip proof stream

[hero78119]

Problem

Concurrent GPU chip proving can return from a worker before its bound CUDA stream is fully drained. This can let the scheduler unbook/reuse resources while async work is still pending, causing intermittent proof corruption such as RootMismatch.

Change

Add a concurrent-mode-only stream synchronization at the end of create_chip_proof_gpu_impl before task resources are released.

Testing

cargo fmt --package ceno_zkvm
cargo check --features gpu --package ceno_zkvm --bin e2e

[hero78119]

Benchmark result from scroll-tech/ceno-reth-benchmark.

• ceno version: 8848c75bb96e7a86088b714c38c74cbde1d1bebb
• benchmark job: failure
• workflow run: https://github.com/scroll-tech/ceno-reth-benchmark/actions/runs/25026562945

[hero78119]

Benchmark result from scroll-tech/ceno-reth-benchmark.

• ceno version: cdbeb78c6790e9d4d6a4eae1d1f9d943e915139f
• benchmark job: failure
• workflow run: https://github.com/scroll-tech/ceno-reth-benchmark/actions/runs/25026771494