feat: support scalar kernel params in runner and test infra

[tetsuo-cpp]

Summary

• Add typed --param format to warpforth-runner: i64:42 (scalar) and i64[]:1,2,3 (array)
• Update conftest.py to parse scalar param declarations and pass them with the typed format
• Add pipeline test (scalar-param.forth) and GPU test (test_scalar_param) for mixed scalar+array kernels

Closes #32

[tetsuo-cpp]

Code Review

Clean PR — scalar param support is well-structured across both the C++ runner and Python test infra. The ParamDecl dataclass is a nice improvement over the raw tuple. A few items to consider:

Actionable

1. Empty value handling in parseParam (warpforth-runner.cpp): Passing --param i64: or --param i64[]: will hit std::stoll on an empty string and throw an uncaught exception. Worth validating that valueStr is non-empty (and that at least one value was parsed for arrays).

2. Pointer-lifetime comment (warpforth-runner.cpp:~214): scalarValues exists so kernelArgs[i] can point at &scalarValues[i] across the kernel launch. This is correct but non-obvious — a brief comment would help future readers.

3. Out-of-bounds output_param (conftest.py:387): The guard output_param < len(decls) silently skips the scalar check when the index is out of bounds, deferring the error to a less obvious place. Consider failing early on out-of-bounds too.

Optional / Nits

• C-style casts in the kernel arg ternary — consistent with the rest of the file but static_cast would be more idiomatic C++17.
• No negative tests for the runner's error paths (bad type prefix, scalar with multiple values, scalar as output). A quick lit test checking for the expected stderr would be nice coverage.
• The MID FileCheck line is sensitive to MLIR attribute ordering — could become flaky if attribute print order changes upstream.

Overall looks good 👍