Adapt Groth16 batch parallelism to pool size

[ValarDragon]

Summary

Adapt multicore Groth16 Miller-loop chunk size to the batch size and Rayon
thread count. The verifier aims for one work item per thread, with chunks
capped at eight proofs. A one-thread pool keeps eight-proof chunks.
Each chunk still receives a nonzero random weight; the same batch pairing
equation and single final exponentiation are used. Smaller chunks expose more
parallelism, especially for batches below the thread count.

The batch benchmark now consumes verification results and includes the multicore
path. This avoids measuring work the optimizer could discard.

Measurement

Local Criterion on a 16-logical-core Mac using the default Rayon pool, valid
one-public-input MiMC proofs:

Proofs	Fixed eight per work item	Adaptive size	Adaptive time	Time reduction
8	6.50 ms	1	2.43 ms	63%
32	6.62 ms	2	4.18 ms	37%
64	7.00 ms	4	6.43 ms	8%

A seven-public-input diagnostic that matches Sapling Spend's input count
measured 7.65 to 3.83 ms (8 proofs), 7.78 to 5.20 ms (32), and 8.37 to
6.79 ms (64) with one-proof chunks. It uses an expanded key and therefore
invalid proofs, but executes the full batch pairing equation. The diagnostic
code is not included. The adaptive choice for a 16-thread pool is one, two,
and four proofs per chunk respectively, so these figures apply directly only
to the eight-proof case.

On Linux 1 (AMD EPYC, eight logical cores), the original eight-proof chunks
measured 12.03, 13.66, and 13.06 ms for batches of 8, 32, and 64. The
adaptive sizes for that pool are one, four, and eight; the pushed revision
measured 6.40, 8.33, and 13.04 ms respectively. The host was idle before the
run.

With Rayon limited to one thread, one-proof chunks regressed from 40.05 to
45.98 ms for 64 proofs. The adaptive implementation retains eight-proof
chunks in that case. With two threads and 64 proofs, eight-proof chunks
measured 22.24 ms versus 23.19 ms for one-proof chunks. With four threads
and 64 proofs, four-proof chunks measured 12.17 ms versus 12.60 ms for
one-proof chunks.

Checks

• cargo test -p zakura-bellman --test mimc batch_verify --locked (valid and
  deliberately invalid multicore batches)
• cargo test -p zakura-sapling-crypto --lib --locked (69 tests)
• cargo check -p zakura-bellman --no-default-features --features groth16 --tests --benches --locked
• cargo fmt --all -- --check
• git diff --check

No public or crate-visible Rust API changes.

[v12-auditor]

Warning

An unexpected error occurred while creating the run. Please try again or use the dashboard.