[CubeCL][P1.1] Implement runtime, device, buffer, transfer, and constructor foundation

[code-tc]

Objective

Implement the reusable CubeCL runtime foundation beneath Array: runtime/device discovery, typed buffers, allocation ownership, host↔device transfer, constructors, and typed unsupported-device/resource errors.

Scope

• CubeCL runtime/client ownership and deterministic backend selection.
• Buffer shape/dtype/layout metadata and allocation lifetime.
• Host upload/download and zero/one/range-style constructors currently supported by the public API.
• Explicit memory accounting hooks and safe cleanup.
• Shared conformance fixtures against host behavior.

Acceptance criteria

• CPU runtime passes constructors and host round-trip fixtures for all supported dtypes/shapes.
• A designated accelerator passes the same fixtures without changing public Array semantics.
• Zero-size, overflow, allocation failure, device unavailable/lost, transfer failure and drop cleanup are typed and tested.
• Device selection never silently claims acceleration after falling back.
• Allocation/transfer evidence exposes bytes, device identity and cleanup without leaking raw data.
• No OpenCL code is removed.

Parent: #35

Tembo execution contract

automation:
  eligible: true
  executor_profile: tembo
  readiness: blocked
  primary_repository: TinyChain-Inc/ha-ndarray
  base_branch: main
  writable_scope:
    - src/cubecl/runtime/**
    - src/cubecl/buffer/**
    - src/cubecl/constructors/**
    - src/platform.rs
    - src/buffer.rs
    - tests/conformance/**
  prerequisites:
    - "#50 designated accelerator validation lane"
    - "#34 accepted HAL/runtime contract"
  concurrency_group: cubecl-runtime-buffer
  required_commands:
    - "cargo fmt --all -- --check"
    - "cargo test --no-default-features --features cubecl"
    - "cargo test --all-features"
  evidence:
    cpu_conformance: required
    accelerator_conformance: required-on-designated-runner
    actual_browser: not-required
    benchmark_report: not-required
    exact_hardware_and_driver_identity: required
  external_actions: none
  draft_pr_only: true
  completion_authority: human
  budget:
    size: large
    max_revisions: 3
    max_wall_minutes: 180