MSVC phase 5: Windows subprocess API in CLI test runner (issue #71)

[dnovick]

Summary

• Adds #ifdef _WIN32 / #else guard to cli_test_runner.hpp so the POSIX fork/pipe/poll/execv/waitpid code compiles only on POSIX; a new Windows path uses CreateProcessA, CreatePipe, PeekNamedPipe/ReadFile, WaitForSingleObject, and GetExitCodeProcess
• Updates io_tests.hpp to replace <sys/stat.h>/<unistd.h> includes and ::stat()/::symlink() calls with #ifdef _WIN32 guarded Windows API equivalents (GetNamedSecurityInfoA for permission checks, CreateSymbolicLinkW for symlink creation with GTEST_SKIP if insufficient privilege)
• Both the strip_trailing_newline parameter and all existing test logic are preserved unchanged

Test plan

• macOS ARM build: cmake --build cmake-build-debug --target safe_crypto_cli_test && ./cmake-build-debug/safe-crypto-cli-test/safe_crypto_cli_test — all tests pass
• All 5 previous MSVC phases (PRs MSVC phase 1: compiler-aware CMake build flags (issue #71) #178–MSVC phase 5: Windows subprocess API in CLI test runner (issue #71) #182) are prereqs; this is the final phase
• Windows CI (when available): verify cli_test_runner.hpp compiles under MSVC without POSIX headers

🤖 Generated with Claude Code

[dnovick]

Blocking review finding after the rebase: in the new Windows runner loop, has_new_data is computed from the current post-drain output size: out.size() + err.size() > (had_data_before ? out.size() + err.size() - 1U : 0U). Once either stream has any data, that expression remains true even when this iteration read zero bytes, so a child process that writes once and then keeps running will bypass both Sleep(1) and the WaitForSingleObject(..., 0) exit check and spin hot. Please compare against a saved byte count from before both drains, e.g. const auto before = out.size() + err.size(); ... const bool has_new_data = out.size() + err.size() > before;.