feat(demo): GPT-2 interactive inference with WarpForth attention

[tetsuo-cpp]

Summary

• Add end-to-end GPT-2 text generation demo using a WarpForth-compiled attention kernel
• Attention kernel uses f32 global memory (F32@/F32!) with f64 shared memory for softmax precision
• PyCUDA wrapper shares PyTorch's CUDA context via autoprimaryctx for zero-copy tensor passing
• Add f32 attention GPU end-to-end test validating reduced-width float memory ops

Closes #46

Test plan

• Verify cmake --build build --target check-warpforth passes
• Run uv run ruff check demo/ gpu_test/ — no lint errors
• On a GPU instance: compile kernel, run gpt2_generate.py, verify coherent text output