Stop the Windows timer thread cooperatively

[kevinushey]

Loading cli starts a timer thread that Windows cleanup currently cancels asynchronously, without waiting for it to terminate. Signal a stop event and join the worker instead, so it returns normally before its resources are released or its DLL is unloaded. Signaling the event wakes the worker immediately, even with a long tick interval.

This removes the architecture-specific cancellation workaround and the unchecked getenv("PROCESSOR_ARCHITECTURE") argument to strcmp(). The Unix worker retains its existing behavior.

Add Windows subprocess tests for missing architecture variables, timer restart, disabling the thread, process exit, and repeated namespace/DLL unloading. Long timer intervals and bounded subprocess timeouts check that shutdown does not wait for the next tick.

Related to #375 (recent report) and #494. The reported Windows crash has not been reproduced locally, so this is a candidate remedy that still needs confirmation on the affected installation.

Validation:

• Built and installed the package on macOS ARM64 with R 4.6.1.
• Full test suite: 2,367 passing assertions, no failures or warnings, and six skips, with inherited NO_COLOR unset. Four skips are the new Windows tests; two require source references.
• Compiled src/thread.c for Windows x64 with MinGW GCC 14.2, POSIX threads, and warnings treated as errors. This checks compilation against Windows APIs; it is not a complete Rtools build or a native Windows runtime test.
• git diff --check passes.