Bound runtime collection launch specialization

[PraneethMerugu]

Summary

• keep collection, keyed-reduction, ordered-fold, and destination-grouping extents as runtime ndrange data
• specialize launches only on their existing 256-lane operation-family workgroup
• add an ordinary CPU science/allocation regression test for the displaced constructor form

Evidence

• live Kaimon: capacity 16→24 = 1 inference event; warmed 300→301 = 1; 24→32 admits KernelAbstractions' one-time dynamic-check mode without extent-specific kernel identity
• helper MethodInstances contain only the operation-family Val{256} family
• warm allocation: baseline 256 B × 7; candidate 256 B × 7
• full CPU: 1,880/1,880
• affected real Metal with scalar indexing disabled: 254/254
• independent review: no findings, including zero-extent behavior
• alternating warm throughput samples found no material CPU or Metal regression

This is the demonstrated C12 LocalMath companion in the canonical Potts PR chain. It adds no cache, policy hierarchy, alternate executor, or backend-specific scientific path.