[codethinki]

The atomic counter alone was not thread safe and not exception safe.

Problems:

• a second `acquire()` returned while the first one was still inside `on`, so the caller used the global state before it existed
• `on` could start while a last `release()` was still inside `off`
• a throw of `on` left the count at 1: `acquired()` was true, and the next `acquire()` did not call `on` again

Fix:

• `acquire()`, `release()` and `force_off()` run under a `std::mutex`
• the count increments after `on` returns
• the count stays atomic, so `acquire_count()` and `acquired()` stay lock free
• `on` and `off` must not call the switch, the mutex is not recursive (doc)

Tests: 5 new tests in `global_switch_test.cpp`. The 3 tests for the problems fail with the old code.

Found while using the switch for the vulkan loader in cth_engine.