Move measurement out of kernel.py for remaining GEAK tasks

[sdubagun-amd]

Moves correctness and measurement ownership out of the agent-editable kernel.py for the two remaining tasks on current main:

• lean_atten_paged
• gemm_a16wfp4

Changes:

• lean_atten_paged: remove correctness, profiling, benchmark, speedup, and CLI code from kernel.py; keep the PyTorch reference and CUDA Event median timing in test_kernel_harness.py.
• gemm_a16wfp4: replace host perf_counter timing with CUDA Event median timing in test_kernel_harness.py.
• Preserve the current main benchmark shape sets and emit canonical ALL_SHAPES indices through GEAK_SHAPES_USED.
• Keep the gemm FP4 capability check in the harness rather than trusting the agent-editable kernel module.
• Emit only GEAK_RESULT_LATENCY_MS; baseline-versus-patched speedup remains the orchestrator responsibility.

The earlier refk_identity and refk_fp8_blockwise_mm changes are omitted because equivalent harness-integrity fixes already landed on main. ff_backward is omitted because the task was removed from main.

Validation:

• MI355X/gfx950 with lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260705.
• lean_atten_paged correctness: 4/4 passed.
• gemm_a16wfp4 correctness: 55/55 passed.
• Full 50-warmup/200-iteration benchmark passed on all 7 lean_atten_paged shapes; geometric mean latency was 0.0883 ms.
• Full 50-warmup/200-iteration benchmark passed on all 55 gemm_a16wfp4 shapes; geometric mean latency was 0.1696 ms.
• Both benchmarks emitted canonical GEAK_SHAPES_USED values and ended with GEAK_RESULT_LATENCY_MS.
• Profile smoke tests passed on 5/5 configs for each task with 1 warmup and 1 measured iteration.
• Python AST parsing, harness ownership assertions, and Docker CLI/import checks passed.

The latency values above are functional benchmark validation results, not controlled leaderboard measurements; GPU clocks were not locked.

[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?