TypedPython kernels: build-time autotuning and a performance baseline in the design #187

What

Add two items to docs/design/typedpython.md §5:

1. Build-time autotuning: tile size, pipeline stages and workgroup size are searched per GPU family (Adreno generations, Mali, Apple GPU generations, desktop) on real devices at build time; the chosen configurations ship in the bundle as a table, and the runtime picks by device. No search on the device.
2. Performance baseline: from the first kernel slice on, measurement tests compare softmax, matmul and attention against llama.cpp (Vulkan, Metal) and Apple MLX on the same device, and record the ratio.

Also point the deferred fusion path at torchnative #68.

Why

The design removes structural ceilings but does not by itself make kernels fast. Without autotuning the first kernels will likely trail hand-tuned code; without a baseline there is no way to claim competitiveness.

Done when

§5 has both subsections, §6 and §7 reference them, no em-dash. Docs only.