perf: batch instance binding in sumcheck stages 5/6

[MatteoMer]

Problem

After computing round polynomials and deriving the challenge, each sumcheck instance's bindChallenge runs as a separate dispatch:

parallelForForce(3, bindRegsVal)           → barrier
parallelForForce(41, bindAllLookupTables)  → barrier
parallelForForce(bytecode_d+1, bindBcRaf)  → barrier
parallelForForce(N+1, bindBooleanity)      → barrier
...

Each bind operation is independent across instances — they all use the same challenge value and operate on disjoint memory. Yet they serialize with barriers between them.

Proposed solution

Fuse all instance bindings into a single dispatch. Instead of N separate parallelForForce calls (one per instance), create a unified dispatch that distributes bind work across all instances:

// Single dispatch: each worker claims an (instance, array_index) pair
parallelForForce(total_bind_arrays, bindAnyArray)  → 1 barrier

Where total_bind_arrays = sum of arrays across all instances (e.g., 3 + 41 + 5 + 4 + ... ≈ 60 arrays in Stage 5).

This gives better load balancing (60 work items across 8 threads vs multiple dispatches of 3-5 items each) and eliminates N-1 barrier cycles.

Files

• src/zkvm/spartan/stage5_prover.zig — bind operations after each round
• src/zkvm/spartan/stage6_prover.zig — bind operations after each round

Interaction

Combines with the "batch instance computation" issue to reduce per-round barriers from 8-9 to exactly 2 (one compute, one bind).