Refactor WitnessBuilder enum into a trait-based extensible architecture

[ocdbytes]

Problem

WitnessBuilder is a monolithic enum (currently 30+ variants) where adding a single new variant requires synchronized changes across 4 files:

1. Enum definition — provekit/common/src/witness/witness_builder.rs (variant + num_witnesses)
2. Solver — provekit/prover/src/witness/witness_builder.rs (witness value computation)
3. Dependency analysis — provekit/common/src/witness/scheduling/dependency.rs (extract_reads / extract_writes)
4. Index remapping — provekit/common/src/witness/scheduling/remapper.rs (remap_builder)

The MSM black-box optimization PR added 10+ new variants (NonNativeEcHint, FakeGLVHint, EcScalarMulHint, SelectWitness, BooleanOr, SignedBitHint,
MultiLimbAddQuotient, MultiLimbSubBorrow, MultiLimbMulModHint, MultiLimbModularInverse, EcDoubleHint, EcAddHint), each requiring all 4 match arms. This is
a pre-existing pattern, but the MSM work significantly exacerbated the maintenance burden — the enum now has ~87 match arms in dependency.rs and ~83 in
remapper.rs.

Key pain points:

• Every new operation touches shared infrastructure files — high merge-conflict risk across teams
• Exhaustive match in 4 files means one forgotten arm = compile error, but the fix is pure boilerplate
• No encapsulation: MSM-specific operations live in common/, visible to all consumers
• num_witnesses() is already a ~30-line match that will only grow

Ref: #310 (comment)