Implement CubeCL reductions and linear-algebra parity

Objective

Complete the numerically sensitive CubeCL tranche for reductions and current linear-algebra operations, including matrix multiplication.

Scope

All existing reductions, aggregation edge cases, matrix multiplication and other current linalg operations. Define deterministic behavior where guaranteed and explicit tolerances where parallel order changes floating-point results.

Acceptance criteria

• CubeCL conformance covers every current reduction and linalg operation supported by OpenCL.
• Empty axes, singleton axes, non-contiguous inputs, mixed supported dtypes, overflow/error cases, and batched shapes are tested where applicable.
• Matmul/reduction outputs match host references within documented tolerances.
• Kernel configuration does not assume NVIDIA-specific warp dimensions.
• Benchmarks record compilation separately from warm execution.

Parent: #33

Tembo-sized implementation issues

• #44 — reductions and numerical conformance
• #45 — matrix multiplication and current linalg operations

#44 and #45 may run concurrently after their shared runtime/elementwise prerequisites.

Tembo program boundary

This parent is a non-executable acceptance rollup. Tembo must be assigned to the child issues, and may not close this issue from a child PR.

automation:
  eligible: false
  executor_profile: tembo
  readiness: rollup
  primary_repository: TinyChain-Inc/ha-ndarray
  base_branch: main
  writable_scope: []
  prerequisites: ["#44", "#45"]
  concurrency_group: cubecl-numerical-rollup
  required_commands: []
  external_actions: none
  draft_pr_only: true
  completion_authority: human

This issue is a non-executable rollup whose implementation is split between child issues #44 and #45. Start by reading those issues and the shared runtime/elementwise prerequisites. Done means satisfying the listed reduction, linear-algebra, matmul, tolerance, portability, and benchmark acceptance criteria.