Phase 2: CPU core-count scaling sweep

[neuromechanist]

Part of epic #84.

Scope

Add a --threads sweep (4/6/8/12/...) to the dimsweep harness for the CPU backends (torch-cpu, numpy, native-fortran) with matched thread budgets, to see whether cores catch the GPU.

Definition of Done

• --threads sweep axis wired to torch (torch.set_num_threads/interop), numpy (OMP/MKL env), and native-Fortran (OMP_NUM_THREADS) with matched budgets per run.
• Results captured per (backend x threads x dims) with per-iteration timing.
• Honest caveats on thread contention documented.
• Real EEG only. NO MOCKS.

[neuromechanist]

Done in PR #89 (squash-merged into the epic branch). Complete cross-platform channel x core sweep, f64+f32, both machines; findings in .context/issue-84/phase2_cpu_scaling.md.