Fix non-memcpy benchmark path saving raw timestamps instead of cycle counts

[ycmath]

In the non-memcpy benchmark path, Program.run calls copy_back_benchmark_data, which returns a (cycle_start, cycle_stop) tuple of absolute timestamps, and binds the tuple to cycle_counts. Consequently:

• perf_cycles.npy stores the two stacked timestamp arrays instead of elapsed cycle counts, and
• the printed "Cycle count" summary is computed from absolute timestamps.

The memcpy path already computes the difference correctly. This changes the call to copy_back_benchmark_cycles, which returns cycle_stop - cycle_start, making both paths consistent.

🤖 Generated with Claude Code

[tbennun]

LGTM!

[ycmath]

Thanks for the review! For the record, the test-csl failure here is a fork-PR limitation rather than this change: the job fetches the SDK with wget "${{ secrets.CSL_SDK_URL }}", and GitHub does not pass repository secrets to workflows triggered by PRs from forks, so the URL comes through empty (log: wget -q -O sdk-1.4.0.tar.gz "" → exit 1; the Linux-deps cache had also expired). test-pytest passes, and main itself is green. Same constraint applies to #72.