Constant memory support for small read-only tables

[tetsuo-cpp]

Summary

Add a way to declare and access CUDA constant memory for small read-only parameters and lookup tables broadcast to all threads.

Motivation

Constant memory is a good fit for:

• Small coefficient tables
• Kernel config vectors reused by every thread
• Broadcast scalars that are awkward as registers-from-params at scale

WarpForth currently has global params and shared scratch only.

Proposed syntax (sketch)

\! const LUT f32[32]
\! const SCALE f64          \ optional: scalar in constant space

Access:

I CELLS LUT + C@     \ or dedicated C@/C! (store only from host)

Host:

• Runner/Python API copies bytes into the constant symbol or uses driver module constants
• Kernel sees read-only pointer / NVVM constant address space

Acceptance criteria

• Header directive for constant buffers (size limits documented, e.g. 64KB device limit)
• Load path lowers to constant address space in PTX
• Host path to populate constant data before launch
• GPU test: all threads read same LUT entry correctly
• Docs: when to use const vs param vs shared

Non-goals

• Runtime-writable constant memory from device
• Exceeding hardware constant capacity (must error clearly)

Related

• Kernel metadata emission
• Python/runner launch APIs (Replace warpforth-runner C++ CLI with Python cuda-python + JSON protocol #50)
• Element-typed declarations