Implement CubeCL array, indexing, transform, and elementwise parity

Objective

Implement the non-reduction CubeCL operation tranche needed for practical Array use while preserving the existing public API and numerical/error semantics.

Scope

Buffer construction and conversion; dtype conversion; broadcasting; unary/binary elementwise operations; slicing, gather, indexing, reshape, transpose and other existing view/transform operations. Reuse shared conformance fixtures rather than copying backend-specific tests.

Acceptance criteria

• Every in-scope public operation supported by OpenCL has a CubeCL conformance case.
• Aliasing, non-contiguous views, bounds errors, zero-sized dimensions, and dtype conversions are covered.
• Host/CubeCL results match within documented dtype tolerances.
• Device allocations and transfers are observable and released after the owning arrays/results are dropped.
• No OpenCL source is changed except shared abstraction/conformance integration.

Parent: #33

Tembo-sized implementation issues

• #41 — runtime, device, buffer, transfer, and constructors
• #42 — broadcasting, elementwise operators, and dtype conversion
• #43 — views, slicing, indexing, gather, reshape, and transpose

Dependency order: #41 → (#42 and #43 in parallel).

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: ["#41", "#42", "#43"]
  concurrency_group: cubecl-general-parity-rollup
  required_commands: []
  external_actions: none
  draft_pr_only: true
  completion_authority: human

Start by reading child issues #41, #42, and #43 in dependency order, then inspect the shared conformance fixtures they identify. The work is done when all in-scope CubeCL operations have conformance coverage, including aliasing, views, bounds, zero-sized dimensions, dtype conversions, tolerances, and allocation-release behavior.