[LuhTonkaYeat]

Adds exec_intersect --- set intersection of two sorted key-value arrays with a
binary operation applied to matched values.

Implementation

• CPU: cpu_intersect.hpp (two-pointer merge).
• OpenCL: cl_intersect.hpp + kernel intersect.cl (naive fallback);
  shared merge module cl_merge.hpp + kernels merge.cl / merge_union.cl.
  Public API, schedule task, registry entries (INT/UINT/FLOAT), C binding
  spla_Exec_intersect.

OpenCL notes

• Merge path: array A is split into tiles, each tile binary-searches its window
  in B; per-tile counts are turned into output offsets with an exclusive scan,
  so the result is deterministic, sorted and written without atomics.
• Shared module: the same primitive is exposed via cl_merge.hpp for other
  algorithms; input and output buffer offsets allow processing CSR row slices
  without copying.
• Union mode: merge of two sorted key-value arrays with equal keys folded by the
  binary operation; uses diagonal merge-path partition (co-rank search) with
  boundary snapping so an equal pair is never split between tiles.
• Hybrid: strongly asymmetric inputs fall back to the binary-search kernel,
  which iterates the smaller array and is faster there.
• Result stored as a coordinate (COO) vector so genuine zero keys/values are
  not confused with empty entries.
• Grid-stride kernel loop keeps large inputs correct despite the grid clamp.
  Null-argument guard for parity with CPU.

Tests

CPU: test_intersect (7 cases).
OpenCL: test_opencl_intersect (22 cases) — types, boundary keys/values, tile
boundaries, exact tile multiples, asymmetric sizes in both directions, large
inputs, negative cases, deferred (scheduled) path, sorted deterministic output.
Merge module: test_merge_union (14 checks) — folding union, intersection,
buffer offsets, tile boundaries, large inputs.