Fix a subtle race condition between prepare_csp_for_next_request() and sweep()

Fix a subtle race condition between prepare_csp_for_next_request() and sweep()

A thread preparing itself for the next client request
could briefly appear to be inactive.

If all other threads were already using more recent files,
the thread could get its files swept away under its feet,
later on causing an 'invalid read of size 1' when the file
name was logged in any_loaded_file_changed().

I've only seen it while stress testing in valgrind while
touching action files in a loop. It's unlikely to have
caused any actual problems in the real world.