Correct the Pixel table: the benchmark token flags were ignored, so the old row was a 19-token prefill mislabelled as 256

Add measured Pixel 8a GPU/CPU rows (litert_lm_main v0.16.0, p256/d256, 3 runs): GPU wins TTFT 8x, prefill 9x, decode too at this size

Backend: GPU support with litert-lm >= 0.16.0 (Android OpenCL + macOS, generation-verified); add measured GPU speed rows; iOS Metal status linked

card: warm-run iPhone decode number (19.7 tok/s)

card: iPhone 17 Pro device results (int4 8/8, ~17 tok/s; int8 loads under EVA entitlements)

LFM2.5-2.6B int8 + int4-b32 .litertlm (litert-lm >= 0.14/0.15), GSM8K 88/83 vs bf16 92

initial commit