UPSTREAM PR #17736: build : move _WIN32_WINNT definition to headers

[loci-dev]

Mirrored from ggml-org/llama.cpp#17736

Previously, cmake was forcing _WIN32_WINNT=0x0A00 for MinGW builds, This caused "macro redefined" warnings with toolchains that define the version.

This also removes the GGML_WIN_VER variable as it is no longer needed.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #417

Overview

PR #417 relocates the _WIN32_WINNT macro definition from CMake build configuration to C/C++ header files (common/common.h and ggml/include/ggml.h). This build system refactoring addresses macro redefinition warnings with MinGW toolchains while maintaining Windows 10 API compatibility.

Performance Impact

Analysis of versions a2b60b7c-9d36-46fe-984d-44f9b3b23fd4 (target) versus 00be72bf-5249-419f-a980-e14c00af57a6 (base) shows no measurable performance changes. All examined functions report 0% variation in both response time and throughput metrics:

• llama_decode: 732255 ns response time, 69 ns throughput (unchanged)
• llama_tokenize: 393910 ns response time, 22 ns throughput (unchanged)
• ggml_backend_graph_compute: 128 ns response time, 22 ns throughput (unchanged)

Power consumption analysis across 16 binaries shows negligible variations within measurement precision (< 0.001%). The largest absolute delta is 0.84 nJ in build.bin.libllama.so, representing a reduction within noise thresholds.

Tokens per Second Impact: None. The inference-critical functions (llama_decode, llama_encode, llama_tokenize) show zero response time or throughput changes, resulting in no impact on token generation rate.

Impacted Binaries: None exhibit meaningful power consumption changes. All variations are sub-nanosecond level.

This PR is a pure build configuration change with no runtime code modifications, confirmed by performance metrics showing complete stability across all measured dimensions.