feat: integrate GPU prover, add basefold modules

[Velaciela]

Description

Performence

• fibonacci $\textrm{fib}(2^{20})$: runs 12M steps (warmup + rerun)
• machine(node-19): AMD EPYC 7702 64-Core, 1 x RTX 3090

# CPU
RUST_LOG=debug cargo run --release --package ceno_zkvm --bin e2e -- --platform=ceno --hints=20 --public-io=3130 --profiling=2   examples/target/riscv32im-ceno-zkvm-elf/release/examples/fibonacci 2>&1 | tee "fib_perf_cpu.log"

# GPU
make enable-gpu
RUST_LOG=debug cargo run --release --features gpu --package ceno_zkvm --bin e2e -- --platform=ceno --hints=20 --public-io=3130 --profiling=2   examples/target/riscv32im-ceno-zkvm-elf/release/examples/fibonacci 2>&1 | tee "fib_perf_gpu.log"
make disable-gpu

• phase1: basefold commit & open

metric	CPU time	GPU time	speedup
ZKVM_create_proof	13.5s	8.0s	1.68x
batch commit to traces	6.07s	1.15s	5.27x
transfer MLEs to host	0	550ms	N/A
transfer pk to device	0	4.53ms	N/A
main_proofs (no GPU yet)	6.04s	6.15s	0.98x
pcs_opening	1.42s	168ms	8.45x

• (WIP) phase2: main & tower prove

metric	CPU time	GPU time	speedup
ZKVM_create_proof	13.6s	6.72s	2.02x
batch commit to traces	6.07s	1.15s	5.27x
transfer MLEs to host	0	550ms (next: eliminate)	N/A
transfer pk to device	0	4.53ms	N/A
main & tower prove	6.04s	4.85s	N/A
-- main (no GPU yet)	-	-	-
-- tower	4.12s	2.72s (next: optimize)	1.51x
pcs_opening	1.42s	168ms	8.45x

Current Status

• Add GPU Prover Hal
• Manually checked that batch_commit and batch_open produce consistent results between CPU and GPU provers
  • cargo test --release test_single_add_instance_e2e -- --nocapture
• API & IO, to pass ceno verifier
• full GPU pipeline for basefold commit -> open
• tower: witness, prove
• main: witness, prove
• perf issues