Move measurement out of kernel.py into the harness for 5 tasks

[sdubagun-amd]

Closes the "kernel.py contains the timer" attack surface in geak_eval tasks where test_kernel_harness.py imported benchmark_config (or run_benchmark/run_correctness/run_profile) from kernel.py. An agent optimising kernel.py could legitimately rewrite the imported timer function and inflate the reported speedup with no actual GPU change (observed previously on refk_identity at 12.07x).

For each of the 5 in-scope tasks the harness now follows the fused_rms_fp8 structural convention:

• refk_identity, refk_fp8_blockwise_mm, ff_backward, lean_atten_paged, gemm_a16wfp4

Per task:

• kernel.py: deleted benchmark_config / check_correctness / run_* / evaluate / main blocks, the in-kernel WARMUP/ITERATIONS module constants, the torch_op wrapper, and any *_pytorch reference.
• test_kernel_harness.py: rewritten as a self-contained script that
  • imports only kernel callables, get_inputs, shape configs, and tolerance constants from kernel.py;
  • carries the moved PyTorch reference and uses it ONLY in run_correctness via torch.testing.assert_close;
  • times only the candidate kernel via torch.cuda.Event median;
  • exposes the standard --correctness / --benchmark / --full-benchmark / --profile modes plus --warmup (default 50) and --iterations (default 200);
  • emits GEAK_SHAPES_USED=[...] and GEAK_RESULT_LATENCY_MS=;
  • does NOT compute a speedup in-process (no GEAK_RESULT_GEOMEAN_SPEEDUP). Speedup is now the orchestrator's job: run the harness twice (unmodified vs patched kernel) and divide.

Notes:

• gemm_a16wfp4: kernel.py was already kernel-only. Harness rewritten to swap perf_counter timing for cuda.Event median and standardise the iteration defaults to 50/200 (was 5/10 and 5/20).
• refk_fp8_blockwise_mm: the new benchmark loop allocates the output c once per cfg (the OLD benchmark_config did c.clone() per iteration, which incorrectly timed an allocation alongside the kernel). The baseline absolute ms will drop slightly on this task as a result.
• lean_atten_paged: --correctness uses CORRECTNESS_CONFIGS (not HARNESS_SHAPES) to match the kernel.py-side convention.

Out of scope (intentionally): orchestrator file-allowlist on patches, which is the sufficient condition to fully defend against an agent that patches the harness file directly.

[irvineoy]

One issue I noticed with the sampled harnesses: GEAK_SHAPES_USED no longer reports the original ALL_SHAPES indices for tasks that downsample.

For example:

• refk_fp8_blockwise_mm/test_kernel_harness.py builds HARNESS_SHAPES from _harness_indices, but later prints GEAK_SHAPES_USED={list(range(len(shapes)))}.
• gemm_a16wfp4/test_kernel_harness.py does the same.

On this PR, the actual sampled source indices are different from the emitted values:

• refk_fp8_blockwise_mm: sampled [0, 1, 2, 4, 5, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 18, 19, 20, 21, 22, 23, 24, 26, 27, 28], but reports [0..24].
• gemm_a16wfp4: sampled [0, 2, 4, 7, 9, 11, 14, 16, 18, 20, 22, 25, 27, 29, 32, 34, 36, 38, 40, 43, 45, 47, 50, 52, 54], but reports [0..24].

If the orchestrator trusts GEAK_SHAPES_USED for attribution or replaying base-vs-patched comparisons, this can point it at the wrong shapes. Could we preserve/pass the selected source indices and print those instead of the local positions?