Add API routine to run code in every thread for updating global flags

On AArchXX architectures, reading a global flag stored in memory in each thread requires a load-acquire operation, which is more expensive than a simple load. For rarely-changed flags, what I'd like to do is have each thread use a regular load, and have an expensive way to perform the load-acquire in every thread all at once when the flag changes. Even better would be to have each thread use a thread-private copy of a flag, and have a way to have each thread load the global value into the private copy when the global changes.

My proposal is a variant of dr_suspend_all_other_threads_ex() that takes in a callback and invokes the callback for every thread. In many use cases it does not need to wait for all threads to reach a synch point: it just sends each a signal and in the handler it invokes the callback (so the callback has to be safe to run from the handler; but I would expect it to normally just do a load-acquire from the global and then store into the private slot). If a synch point is needed, we could have a callback that's invoked from the suspension point of dr_suspend_all_other_threads_ex().

My use case is #3995 where I want to change my drbbdup (#4134) case value to swap between instrumentation modes. The idea is to avoid the cost of reading the global flag holding the mode at the top of every single block.