Configurable GPU architecture target (sm_XX) for warpforthc / pipeline

[tetsuo-cpp]

Summary

Replace the hard-coded NVVM chip sm_70 with a user-selectable architecture on warpforthc, warpforth-opt pipeline options, and the test harness.

Motivation

buildWarpForthPipeline currently sets:

nvvmOptions.chip = "sm_70"; // Volta

That freezes codegen at Volta. Newer features (tensor cores shapes, async copy, TMA-oriented paths, better libdevice matching) need sm_80 / sm_89 / sm_90 (and future chips). Hard-coding also blocks honest performance work for #11 (MMA) and modern hardware used on Vast.ai.

Proposed CLI

warpforthc kernel.forth -o kernel.ptx --arch sm_80
warpforthc kernel.forth --arch=sm_90

Defaults: keep sm_70 for compatibility, or document a new default if tests allow.

Also expose the same option on:

• warpforth-opt --warpforth-pipeline (pass pipeline options)
• GPU test harness / runner so remote GPUs compile for a compatible arch

Acceptance criteria

• --arch (or equivalent) on warpforthc sets GpuNVVMAttachTarget chip
• Invalid arch names produce a clear error
• Document supported/tested arches in README
• CI still green on default arch
• GPU tests can pass arch through to the compiler

Related

• Tensor core / MMA (Tensor core / MMA intrinsics #11)
• Optimization pipeline expansion
• Runner redesign (Replace warpforth-runner C++ CLI with Python cuda-python + JSON protocol #50)