micro-gpt across the abstraction stack

Headline findings

• NumPy spends ~96% of wall-clock on dispatch/typecheck overhead, only ~4% on the actual matmul kernel.
• MLX-GPU (1,865 tok/sec) loses to pure Python — GPU launch overhead is the wrong shape for 4K MACs/token.
• WebAssembly hits 25× the FPGA in regular Chrome, ~35% of LUT-optimized native C+NEON.
• The model has at most 8 effective bits of precision. Per-tensor int8 shows no measurable degradation.
• Same WASM binary, different V8 builds = ~50% throughput delta. Chrome 145 vs Electron 41's Chromium 146.