fix(jit): add driver-only mode support for LaunchConfig

[m96-chan]

Summary

Fix the test-driver-only-windows CI job failure by adding driver-only mode support to native/jit/kernel.hpp.

Problem

The CI job failed with errors like:

error C3646: 'grid': unknown override specifier
error C2061: syntax error: identifier 'dim3'

This occurred because dim3 and cudaStream_t are CUDA Runtime types, but driver-only mode (PYGPUKIT_DRIVER_ONLY=ON) only includes the CUDA Driver API headers.

Solution

• Add a portable Dim3 struct that's used in driver-only mode
• Use StreamHandle from stream.hpp instead of cudaStream_t
• Follow the existing pattern used in stream.hpp

#ifdef PYGPUKIT_DRIVER_ONLY
struct Dim3 {
    unsigned int x, y, z;
    Dim3(unsigned int x_ = 1, unsigned int y_ = 1, unsigned int z_ = 1)
        : x(x_), y(y_), z(z_) {}
};
#else
#include <cuda_runtime.h>
using Dim3 = dim3;
#endif

Test plan

• Normal build succeeds locally
• CI test-driver-only-windows job passes

🤖 Generated with Claude Code

[m96-chan]

Summary of Changes

This PR now includes three fixes:

1. Driver-Only Mode Support (kernel.hpp)

• Added portable Dim3 struct for PYGPUKIT_DRIVER_ONLY builds
• Changed cudaStream_t to StreamHandle (already defined in stream.hpp)
• Verified: Local driver-only build succeeds (sm_80/86)

2. README False Claims Fix

• Removed misleading "no mandatory external SDKs" claim
• Added explicit note: "PyGPUkit currently requires CUDA drivers and NVRTC"
• Clarified: "NOT a PyTorch/CuPy replacement"

3. Condensed Roadmap

• Replaced 37-line checklist with compact table for v0.1-v0.2.3
• Keeps future versions (v0.2.4+) as detailed checklists

| Version   | Highlights |
|-----------|------------|
| v0.1      | GPUArray, NVRTC JIT, add/mul/matmul, wheels |
| v0.2.0    | Rust scheduler, memory pool, kernel cache, 106 tests |
| v0.2.1    | API stabilization, error propagation |
| v0.2.2    | Ampere SGEMM (cp.async, float4), 18 TFLOPS FP32 |
| v0.2.3    | TF32 TensorCore (PTX mma.sync), 27.5 TFLOPS |