bench: a performance model, to estimate decode and encode times on devices we cannot measure #159

Why

Measuring on a phone needs the user's phone, unattended, with the keychain unlocked. That does not scale, and each release should still say roughly what it costs on other devices.

The numbers already measured on the M3 Max and the iPhone 17 (release 0.11.0, through the C API, same app, same pictures) fit a two-term model: a fixed cost plus a cost per pixel. The iPhone's per-pixel cost is about 5 times the M3 Max's, near their memory bandwidth ratio (400 against 68 GB/s) and far from their GPU core ratio.

What

• scripts/perf_model.py, standard library only: fits time = fixed + per_pixel x pixels for each way (decode to planes, decode to RGB, encode) on the M3 Max, compares the fit with the iPhone 17's measurements, and prints estimates for other chips from their memory bandwidth.
• A table in BENCHMARKS.md, kept apart from the measurements and marked as estimated, with the basis and date of every number.
• What the model cannot give stays measured: the cost of decoding at a stream's pace (the GPU's power states), which is where bench: decode latency inside a client's display loop #81 comes in.

Done when

The script runs from a clean checkout, its fit and its check against the iPhone are printed, and BENCHMARKS.md has the estimated table with its limits stated.