cxa_thread_call_dtors(): restore limit on the number of destructors called

[kib]

Summary

If the exiting thread has a destructor registered that keeps
re-registering itself, the function loops forever. Implement
approximation for the number of passes over the desctructor list as it
was done before 9d26b82826d99.  Maintain the count of destructors
registered, and break out of the loop if we called more destructors then
the memoized count times CXA_DTORS_ITERATIONS.

__cxa_thread_call_dtors(3): bump number of iterations

This allows more breezing room for somewhat weird code that registers
thread-local destructors at destruction time.  At the same time, we
still limit the amount of iterations, making this aspect of the thread
termination time-bound.

[kib]

Bump the iteration limit. Request review.

[markj]

Seems ok, but I'm not sure if it's sufficient for the LLVM test case that @aokblast was looking at.

lib/libc/stdlib/cxa_thread_atexit_impl.c

129

cnt is uninitialized.

[kib]

Initialize cnt

[aokblast]

In D56053#1284052, @markj wrote:

Seems ok, but I'm not sure if it's sufficient for the LLVM test case that @aokblast was looking at.

It works on my test but a problem comes from my previous comment in D56022

ignore my incomplete comment

lib/libc/stdlib/cxa_thread_atexit_impl.c

126

Can we just