bench: NG vs Fortran runtime harness (#59)

[neuromechanist]

Part of #59. Adds benchmarks/benchmark_runtime.py, a repeatable runtime benchmark for AMICATorchNG vs the Fortran binary on the real sample EEG.

What it does

• Matched config (same iters, lrate, Newton schedule, num_mix, pdftype) pulled from sample_params.json / input.param, so both backends do equivalent work (verified by near-identical final LL).
• Sweeps an iteration x thread-count grid; reports total wall-clock (mean/std) and a fixed-overhead-free marginal per-iteration slope (cancels data load / process spawn / I/O).
• Environment banner flags that the bundled amica15mac is x86_64 and runs under Rosetta 2 on arm64, so local ratios flatter NG.
• Writes JSON + a markdown report (dir gitignored). No silent failures: a nonzero Fortran exit raises.

Measured result (this host: M4 Pro, Fortran emulated)

NG misses the 2-3x criterion on the CPU float64 parity path: ~3.2x single-thread, ~7.3x at 8 threads because NG barely parallelizes on CPU (219->204 ms/it) while Fortran's OpenMP scales ~2.4x. Full numbers + profile posted to #59; hot-spot follow-up filed separately.

Docs-adjacent tooling; no library code touched.