Feat: less mmcs_verify_batch

[kunxian-xia]

Summary

This is a reflection of optimization scroll-tech/gkr-backend#24 in gkr-backend to the recursion crate.

The key idea is to reduce the number of mmcs_verify_batch invocated for the integrity of input opened values.

• The number of rows in the FriReducedOpeningAir is reduced by half.
• The number of rows in the NativePoseidon2Air for VERIFY_BATCH is reduced by half.

Performance

The number of cells in the recursion prover before and after this PR is depicted here.

The cmd to run the test is

RUST_BACKTRACE=1 RUST_LOG=info,openvm_cuda_common=warn \
  cargo run --release --features "gpu,bench-metrics"  --package ceno_recursion --bin e2e_aggregate \
  -- --platform=ceno --max-cycle-per-shard=1600 \
  examples/target/riscv32im-ceno-zkvm-elf/release/examples/keccak_syscall 2>&1 | tee keccak.log

chip	before	after
total_trace_cells	703_991_274	613_158_378
NativePoseidon2	123_469_824	61_734_912
FriReducedOpening	58_195_968	29_097_984

[hero78119]

A quick pass and LGTM!