[Streaming][S2] Execute Array-block streams with reusable fused CubeCL plans

[code-tc]

Objective

Run compatible Array block streams through reusable fused CubeCL plans with bounded host/device memory and observable transfer/compute scheduling.

Acceptance criteria

• Compatible blocks reuse a compiled plan rather than recompiling per block.
• A test stream larger than device memory completes within an explicitly asserted peak-memory budget.
• In-flight host blocks, device allocations and compilation work remain bounded under a slow consumer.
• Transfer/compute overlap is enabled only when supported and cannot reorder results.
• Cancellation, device loss, block failure and consumer drop release buffers and produce typed partial/terminal evidence.
• Results match sequential unfused host execution.
• Benchmark report separates compilation, transfer and execution time.

Parent: #38

Tembo execution contract

automation:
  eligible: true
  executor_profile: tembo
  readiness: blocked
  primary_repository: TinyChain-Inc/ha-ndarray
  base_branch: main
  writable_scope:
    - src/stream/**
    - src/fusion/**
    - src/cubecl/**
    - tests/stream/**
    - benches/**
  prerequisites:
    - "#50 designated accelerator validation lane"
    - "#47"
    - "#48"
  concurrency_group: array-stream-cubecl
  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: 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