[CubeCL][P2.1] Implement reduction kernels and numerical conformance

[code-tc]

Objective

Implement the current reduction/aggregation surface on CubeCL with explicit empty-axis, accumulation, ordering and numerical-tolerance contracts.

Acceptance criteria

• Inventory identifies every public reduction and its supported dtype/axis behavior.
• Full, axis, multi-axis if supported, singleton, empty, non-contiguous and large inputs are covered.
• Accumulator dtype, overflow, NaN and infinity behavior is documented and fixture-tested.
• Deterministic guarantees are stated; nondeterministic floating ordering uses bounded per-operation tolerances.
• CPU and designated accelerator results match the host oracle within those contracts.
• Cold/warm performance and temporary-memory use are recorded.

Parent: #36

Tembo execution contract

automation:
  eligible: true
  executor_profile: tembo
  readiness: blocked
  primary_repository: TinyChain-Inc/ha-ndarray
  base_branch: main
  writable_scope:
    - src/cubecl/reduce/**
    - src/ops.rs
    - tests/conformance/**
    - benches/**
  prerequisites:
    - "#50 designated accelerator validation lane"
    - "#41"
    - "#42"
  concurrency_group: cubecl-reductions
  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