TypedPython: commit to GPU kernels and CPU SIMD, record the kernel backend decision #185

What

Promote GPU kernels (and CPU SIMD for @compiled) in TypedPython from an exploration to a commitment, and rewrite docs/design/typedpython.md §5 with the backend decision from the two investigations:

• Frontend and IR are ours (valid Python, runs on CPython); the IR is designed to map onto MLIR.
• Emitters per target: SPIR-V (Vulkan), MSL emitted directly (Apple, so simdgroup_matrix and Metal 4 tensor ops stay reachable), WGSL (web), PTX (desktop CUDA), C with vector extensions / NEON (CPU SIMD).
• The MLIR Python wheel is an optional build-time helper (SPIR-V/PTX serialisation, validation, cross-check), run in a separate 3.14t process until a 3.15t wheel exists.
• Runtime is torchnative's device layer (ash Vulkan, candle-metal); wgpu only for the web.
• Tile-first kernel model, SIMT as the fallback.
• TileLang/TVM, CubeCL, IREE, xDSL: references, not dependencies.

Why

The maintainer wants TypedPython to stand against Mojo; INTENT §1.4 still calls GPU kernels an exploration. The investigations are recorded under docs/investigations/.

Done when

INTENT §1.4 states the commitment with the maintainer's words; design §5, §6, §7, §8 are consistent with the decision; both investigation records are in docs/investigations/; no em-dash in changed files. Docs only, no code.