eval: hand kernel a fresh input each timed iteration (close result-replay in scored path)

[nikhilbarhate99]

Problem

examples/eval.py::_run_single_benchmark runs the scored (benchmark / leaderboard) path with recheck=False: it calls generate_input() once, checks correctness once, then invokes custom_kernel(data) for all max_repeats timed iterations on the same data object (same id(), same .data_ptr()).

This leaves the dominant "result replay" hack open in the ranked path: a submission can cache the output on the first call (keyed on object identity / .data_ptr() / ._version) and return it with zero compute for every timed iteration, reporting an enormous false speedup. (Test mode already uses recheck=True, but the ranked timing does not.)

Fix

In the non-recheck timed loop, hand the kernel a fresh tensor clone each iteration (new object + new data_ptr()). The clone happens outside the timed region (before start_event.record()), so the measured kernel time is unaffected. The two most recent inputs are retained so the caching allocator can't immediately recycle a freed data_ptr.

This defeats identity / ._version / pointer-based result replay in the scored path at negligible (untimed) cost.

A stronger alternative (already supported by the harness) is to default recheck=True for ranked, which additionally re-generates with a fresh seed and re-checks correctness every iteration; the clone approach is the minimal, low-overhead fix that closes the replay vector specifically.

Marking as draft for discussion.