Docs: TypedPython kernel autotuning and performance baseline (#187)

[b-re-w]

Closes #187

Docs only. docs/design/typedpython.md:

• §5.7 build-time autotuning: per GPU family on real devices, configuration table shipped in the bundle, no on-device search.
• §5.8 performance baseline: softmax, matmul, attention against llama.cpp (Vulkan, Metal) and MLX on the same device from the first slice; target ratio set by the maintainer after the first measurement.
• §5.9 fusion deferred to torchnative Stubs: extension properties on a class whose stub name is taken in its module are dropped (Icons.Default.Add) #68.
• §6 steps 4 and 5, §7 risks updated; Taichi moves to §5.10.