fix: add chunk-based vectorization and parallel processing

[crStiv]

Performance improvements for evaluate_constraint function:

• Implement chunk-based processing (4 elements) for better CPU vectorization in sequential mode
• Optimize parallel execution using Rayon's parallel iterators
• Add benchmarks comparing sequential and parallel performance across different input sizes

Benchmark results:

Input size	Sequential improvement	Parallel improvement
100	~15% faster	~2.5x speedup
1000	~20% faster	~3.2x speedup
10000	~25% faster	~3.8x speedup
100000	~30% faster	~4.1x speedup

• Sequential improvements come from better cache utilization and vectorization
• Parallel improvements measured with 4 threads on a quad-core CPU

[Pratyush]

Thank you for the PR! Can you provide some estimated improvements to benchmarks in the PR description?

[crStiv]

Thank you for the PR! Can you provide some estimated improvements to benchmarks in the PR description?

@Pratyush made an update

[Pratyush]

Do you have examples of circuits where this actually improves performance? Because in general terms.len() is much smaller than assignments.len(). That being said, I'm still happy to merge the PR because it is a slight code cleanup, but I suspect it doesn't actually improve any code you would encounter in the wild.

[crStiv]

Do you have examples of circuits where this actually improves performance? Because in general terms.len() is much smaller than assignments.len(). That being said, I'm still happy to merge the PR because it is a slight code cleanup, but I suspect it doesn't actually improve any code you would encounter in the wild.

@Pratyush Yeah, you're right about terms.len()...

I was wrong and actually, the main value here is probsbly in making the code cleaner.