bench: contribute a numeric_cliffs.py datapoint for your chip

[sbryngelson]

Companion to the #137 roofline drive, but for correctness ceilings instead of throughput. bench/numeric_cliffs.py measures the magnitudes at which fp16 math on the engine silently returns the wrong answer -- the three cliffs ANEForge has characterized (see the module docstring and docs/cross-chip.md):

1. matmul_saturation -- a matmul output flips finite->inf near fp16_max/2 (~32752), for every contraction size K (matmul saturates at fp16_max/2 (~32752); two transpose-fed cases return inf even below that #115).
2. slice_saturation -- a width-offset slice routing through the pre-A16 Q.4 x16 crop-DMA that clamps |value| > 4094 to +/-inf. Documented EXACT on A16/M5, so a modern part should report no cliff.
3. reduce_exactness -- a reshape->reduce_sum of small integers stops being bit-exact crossing the fp16 integer-spacing boundary at 2048.

Each cliff is potentially silicon-dependent, and that is exactly the cross-hardware question the roofline work asks: does your chip's matmul cliff sit where the M2/M5 cliff does? Is the slice cliff present (pre-A16) or gone (A16/M5)?

How to contribute (2-minute run, you get credited):

PYTHONPATH=. python3 bench/numeric_cliffs.py

Paste the JSON it prints as a comment here, along with your chip (e.g. "Apple M3 Pro, 18-core GPU"). No code required. If you want, bench/_machine.py (python3 -c "import bench._machine as m, json; print(json.dumps(m.fingerprint()))") prints a machine fingerprint you can paste alongside.

Good first issue for anyone with an Apple Silicon Mac -- the more chip generations we get, the sharper the cross-silicon cliff map.