Add Windows CMake build + workaround NVIDIA OpenCL #include bug (enables Blackwell sm_120)

[Dalailalama]

Summary

This PR adds three independent but related improvements that together enable Vina-GPU 2.1 on Windows + NVIDIA Blackwell GPUs (RTX 50-series, sm_120):

1. Native Windows CMake build system (CMakeLists.txt) — upstream ships only a Linux Makefile and a README describing manual Visual Studio project setup. This commit adds a CMake build that works with VS2022 + MSVC 19.44 + Boost 1.83 + CUDA 13.x out of the box. CMake 4.x compatible.

2. Workaround for NVIDIA OpenCL #include driver bug (code_head.cl) — NVIDIA's OpenCL driver (595.97+ on Linux, 596.36+ on Windows; confirmed on Blackwell sm_120) silently fails clBuildProgram with CL_OUT_OF_HOST_MEMORY (err=-6) and zero log bytes when kernel source contains a #include directive. This affects every Vina-GPU 2.1 user on a current NVIDIA driver because code_head.cl begins with #include "kernel2.h". Workaround: inline kernel2.h content directly. AMD/Intel drivers continue to work as before.

   Reference: https://forums.developer.nvidia.com/t/bug-opencl-include-directive-causes-cl-out-of-host-memory-on-driver-595-97/366719

3. Verbose OpenCL build log on failure (wrapcl.cpp) — existing error handling has a buffer-overrun bug (malloc(logsize) without +1 for null terminator) and stdio buffering issues (no fflush(stdout) before exit(-1)). On platforms where the OpenCL driver returns logsize=0, the existing code prints uninitialized memory like log:C. Fixed by adding null termination, fflush at every print site, explicitly printing the clBuildProgram error code, and mirroring the log to opencl_build_log.txt as backup against pipe-buffering edge cases.

Validation

Tested with a representative docking benchmark (drug-like ligand + GPCR receptor from the PDB) on:

• Hardware: NVIDIA GeForce RTX 5050 Laptop GPU (Blackwell, sm_120, 8 GB VRAM)
• Driver: NVIDIA 596.36 (CUDA 13.2 OpenCL ICD)
• OS: Windows 11
• Toolchain: VS2022 Build Tools, MSVC 19.44.35207, Boost 1.83.0, CMake 4.3.1

Result:

Run	Wall time	Mode-1 score
Cold (kernel compile + dock)	14.98 sec	within expected range
Warm (kernels cached)	2.06 sec	deterministic (same score across runs)

GPU utilization observed at 100% during dock (nvidia-smi sample, 126 MiB VRAM, 51W). Output PDBQT is in standard AutoDock Vina format.

Impact

To the best of my knowledge this is the first documented successful Vina-GPU 2.1 run on an RTX 50-series (Blackwell) GPU. Prior to these patches, on the same hardware:

• Kernel compilation failed silently (CL_OUT_OF_HOST_MEMORY, no log) due to the NVIDIA #include driver bug.
• Windows users had no upstream-supported build system at all.
• Debugging was difficult because the build log was truncated/garbled.

All three commits are independent — each can be cherry-picked separately if preferred.

Notes

• All three commits keep changes minimal and focused — no algorithmic changes to docking.
• The CMake build also works on Linux/macOS with appropriate -DBOOST_ROOT=... and -DCUDA_ROOT=... (not tested by me, but the build script is platform-agnostic).
• LF->CRLF line-ending warning when committing code_head.cl on Windows is git's standard autocrlf normalization (harmless; content stored as LF in the repo).