perf: Stage 6 sumcheck micro-optimizations

[MatteoMer]

Summary

• Lift RaPolynomial tag dispatch out of inner loops: Both RamRaVirtualProver and LookupsRaVirtualProver mapFn bodies now dispatch on the first ra_poly's active tag once via inline else, generating specialized loop bodies per variant. Inner loops use @field(..., @tagName(tag)).getBoundCoeff() instead of the tagged union's switch-based dispatch, eliminating per-access tag checks in release mode.
• Add has_nulls branchless fast path: RaPolynomial compressed rounds (Round1/2/3) now track a has_nulls: bool flag, computed during initRound1 and propagated through bind transitions. When false (common for instruction lookups), getBoundCoeff skips optional index checks entirely.

Context: flamegraph analysis of sha256_2048 showed Stage 6 at 7.2% of total prover time. These are diminishing-returns micro-optimizations — the implementation already closely matches Jolt's approach.

Benchmark (sha256_2048, 10 runs, ReleaseFast, Apple Silicon)

Metric	Before (ms)	After (ms)	Delta	Change
Mean	2814	2672	-142	-5.1%
Median	2759	2651	-108	-3.9%
P25	2690	2607	-83	-3.1%
Min	2582	2540	-43	-1.6%
Stdev	188	105

Improvement is larger than the conservative 0.5–1% estimate — the inline else dispatch likely enables LLVM to further optimize the inner loops (inlining, register allocation) once the variant is statically known. Reduced stdev (188→105) suggests fewer branch mispredictions.

Test plan

• zig build test — all 539 tests pass
• zig build -Doptimize=ReleaseFast — clean build
• Before/after benchmark on sha256_2048 (10 runs each)

🤖 Generated with Claude Code