ci(codspeed): drop simulation (cachegrind); memory is the always-on baseline

[FBumann]

TODO (human): why drop simulation — your words.

Note

The following content was generated by AI (Claude).

Drop the CodSpeed simulation (cachegrind) instrument. Instruction counts
under-weight the numpy/scipy/sparse work this fork targets, and the cachegrind
pass was ~11 min — by far the slowest instrument. This project doesn't do
micro-CPU tuning, so simulation's deterministic-sub-1%-CPU value doesn't earn it.

Dropping it also removes the whole reason codspeed.yml existed — memory and
simulation shared one run only to avoid a same-environment overwrite.

Result

• codspeed-memory.yml — memory on master + every PR + dispatch (the
  always-on baseline; no more label gate). ~2 min.
• codspeed-macro.yml — walltime on master + dispatch + labeled PRs (unchanged).
• codspeed.yml — deleted.

memory (ubuntu) and walltime (macro) are separate environments, so their uploads
don't clash.

What runs when

• Every PR → memory (~2 min)
• Master / dispatch → memory + walltime
• trigger:benchmark-labeled PR → memory + walltime

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will not alter performance

✅ 79 untouched benchmarks
⏩ 495 skipped benchmarks¹

---

_{Comparing ci/codspeed-drop-simulation (9af13c7) with master (f37c8be)}

Footnotes

1. 495 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩