[Profile] further improvements for Windows-only

[vtjnash]

• Instead of preventing profiling entirely (which might deadlock if DllInit accesses Julia in any way), use a suggestion by @gbaraldi to abort the profile if it takes too long (100ms), allowing the stopped thread to release any locks and free up both threads to continue all. Many, many other projects have encountered this same bug, including dotNet (closed as won't fix) and Firefox (uses a single stepping debugger to find the memory address of the offending locks), but those seem difficult responses for us.
• Define correct ordering for taking profile-related locks
• Profile all threads instead of just main thread
• Refactor needsSymRefreshModuleList to use a modern API for doing so reliably. The API does not appear to implement the documented behaviors, but it seems to do what we need.
• Make stack unwinding on Windows thread-safe so that using multiple threads that throw exceptions won't cause memory corruption anymore
• Use condition variables for correctly controlling profiler start/stop, instead of volatile

[IanButterworth]

Profile all threads instead of just main thread

Great! Worth a NEWS entry and the docs for Profile.print have a note that needs removing

[vtjnash]

Sure, do you want to do that and also fix jl_profile_task with similar code?

[IanButterworth]

No I'm good thanks, go ahead.