Zf/zkcuda_service mode

[zhiyong1997]

We're currently calling mpi executor whenever we receive a request to prove a circuit. This works but the circuit needs to be reloaded each time we call the prove, which is very costly. This PR targets at making the proving a service so that the circuit can be pre-loaded.

• Basic Framework.
• Testing and Debug (The compatibility of MPI and the internal multi-threading of axum package is a big issue now).

It is still too costly to call the proving service kernel-by-kernel. To remedy this, we'd like to group the witness together and pass it to the service at once. Note that this breaks the current zkcuda framework a little bit, so we're changing the zkcuda framework first to allow more flexible proving system.

• Zkcuda framework change to allow flexible proving strategy.
• Huge refactor, move almost all the computation to the server.
• Minimize the memory usage across mpi processes.
• Add batch opening and verification for PCS to boost the performance.

Known Issues

1. CI occasionally fails due to rust toolchain setup.
2. When there are not enough physical cores, mpi overscribe will be enabled, but the proving will fail. MPI apis do not seem to behave as expected in the overscription mode.

[zhiyong1997]

parallelized zkcuda for keccak pass.