bench(ops): add matmul micro-benchmarks (dense vs sparse operand)

[FBumann]

We are not yet tracking sparse matmul in the ops. To optimize it, im adding a baseline benchmark here.

Sorry for the churn. I wanted to keep it in the kvl benchmark, but It regressed with matmul, so decided to keep the existing one.

Note

The following content was generated by AI (Claude Code), prompted and reviewed by @FBumann.

The suite has no benchmark exercising @/dot: the kvl_cycles pattern — written to show what a sparse-aware matmul would win (#748) — builds its constraint via the expanded (flow * C).sum("branch"), which bypasses __matmul__ entirely, and ops.py has no contraction op. A sparse-aware kernel (#867) would land invisible to CI.

An earlier revision of this PR switched kvl_cycles itself to the @ operator, but that surfaced as a flat +31% memory "regression" across all severities: on the current kernel, Variable.__matmul__ goes through to_linexpr() (ones coefficients) plus a full-size multiply, allocating an extra (time × branch × cycle) float64 temporary that the expansion avoids. Real information about the @ idiom's current cost, but the wrong place to change measurement history — so instead this adds a dedicated matmul op group to ops.py, following the suite's own op-vs-pattern philosophy ("an op benchmark says which path got heavier"):

• expr_matmul_dense — fully nonzero (1000 × 100) operand; stays on the dense kernel, guards it.
• expr_matmul_sparse — incidence-shaped operand at 0.3% density (the @/dot against a sparse matrix densifies the result to full _term #748 KVL case, ~3 of 1000 entries per column); flat today, collapses under a sparse-aware kernel.

New ops carry no CodSpeed history, so no existing baseline moves and nothing is flagged. kvl_cycles is untouched. Landing this before #867 gives that PR a proper baseline: its CodSpeed run should show expr_matmul_sparse dropping sharply while expr_matmul_dense stays put.

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will not alter performance

✅ 173 untouched benchmarks
🆕 2 new benchmarks
⏩ 173 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
🆕	Memory	test_op[expr_matmul_dense]	N/A	39.7 MB	N/A
🆕	Memory	test_op[expr_matmul_sparse]	N/A	39.7 MB	N/A

---

_{Comparing bench/kvl-matmul-idiom (d38c0a4) with master (2afe5d0)}

Footnotes

1. 173 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. ↩

[FBumann]

I will merge this as soon as tests finish