Introduce backend abstraction and minimal CubeCL backend

Objective

Add a backend-neutral execution boundary and an opt-in cubecl feature, proving one representative elementwise operation on CPU and a supported accelerator without changing the public Array API or removing OpenCL.

Scope

• Define backend/device/buffer/dispatch/error traits owned by ha-ndarray.
• Add CubeCL dependency and feature selection with a documented compatibility pin.
• Implement allocation, host upload/download, device selection, and one unary and one binary elementwise kernel.
• Preserve the host and OpenCL paths during migration.
• Add capability discovery; unsupported targets must fail with a typed error, never silently change semantics.

Acceptance criteria

• --no-default-features --features cubecl builds and tests on a documented CPU runtime.
• A supported GPU/WebGPU target passes the same representative operation fixtures.
• Existing host/OpenCL tests remain green.
• Shape, dtype, broadcasting, and error behavior match the current API.
• Backend selection and fallback rules are documented and covered by tests.
• The PR records the CubeCL version and current supported runtime matrix.

Non-goals

Full operator parity, fusion, browser packaging, OpenCL removal, and replacement of Rayon.

Parent: #33

Tembo validation additions

Before implementation, the PR must add an architecture decision record identifying the exact CubeCL version/revision, enabled runtimes/features, supported Rust version, public HAL traits, backend-selection/fallback rules, typed capability errors, and compatibility/update policy. CubeCL is the kernel/runtime substrate; ha-ndarray owns its public HAL, execution IR, fusion legality, and observable semantics.

The validation report must distinguish CPU-runtime proof from actual accelerator proof and record CubeCL, Rust, OS, device, driver and runtime identities. The agent may not claim accelerator support based only on compilation or CPU emulation. A documented accelerator smoke run is sufficient for this architectural vertical slice; #50 subsequently makes that validation reproducible and is not a reverse prerequisite for #34.

automation:
  eligible: true
  executor_profile: tembo
  readiness: ready
  primary_repository: TinyChain-Inc/ha-ndarray
  base_branch: main
  writable_scope: [Cargo.toml, Cargo.lock, src/platform.rs, src/buffer.rs, src/cubecl/**, tests/**, docs/**]
  prerequisites: []
  concurrency_group: cubecl-hal-foundation
  required_commands:
    - "cargo fmt --all -- --check"
    - "cargo test --no-default-features --features cubecl"
    - "cargo test --all-features"
  evidence:
    architecture_decision: required
    cpu_conformance: required
    designated_accelerator_conformance: provisional-smoke-required
    reproducible_accelerator_lane: "follow-up #50"
    exact_environment_identity: required
  external_actions: none
  draft_pr_only: true
  completion_authority: human-architecture-review
  budget: {size: large, max_revisions: 3, max_wall_minutes: 180}

Start with the architecture requirements and inspect Cargo.toml, Cargo.lock, src/platform.rs, src/buffer.rs, src/cubecl/, tests/, and docs/**. Run the specified fmt and feature-gated test commands to establish the baseline. Done requires the documented CubeCL decision, public HAL behavior, CPU and designated accelerator evidence, preserved host/OpenCL tests, and typed capability and fallback behavior.