Extract shared search_core for MULTI_CTA CAGRA search

The single- and multi-partition MULTI_CTA kernels duplicated the entire
search loop (~100 lines), differing only in indexing and per-partition
data plumbing. Factor the body into a force-inlined search_core() — mirroring
the SINGLE_CTA path — parameterized by the buffer/traversal graph degree,
optional seeds/source-indices/stats, and a BitsetT template arg. The two
kernels become thin wrappers; multi-partition is a strict generalization
(gridDim.z==1 collapses to the single-partition case), so single-partition
behavior is unchanged and the mp path regains the _CLK_BREAKDOWN
instrumentation that the copy had dropped.

Author: jamxia155