Unfortunately, there's not much we can do about it: if dlsym() fails - which is the case here either because read() fails with EBADF, or because the file descriptor now points to another stream (i.e. not libgcc), the libc aborts (here upon pthread_exit(), not PyThread_delete_key()):
"""
libgcc_s.so.1 must be installed for pthread_cancel to work

Program received signal SIGABRT, Aborted.
[Switching to Thread 0xb7b0eb70 (LWP 17152)]
0xb7fe1424 in __kernel_vsyscall ()
(gdb) bt
#0  0xb7fe1424 in __kernel_vsyscall ()
#1  0xb7e4e941 in *__GI_raise (sig=6) at ../nptl/sysdeps/unix/sysv/linux/raise.c:64
#2  0xb7e51d72 in *__GI_abort () at abort.c:92
#3  0xb7e8ae15 in __libc_message (do_abort=1, fmt=0xb7f606f5 "%s") at ../sysdeps/unix/sysv/linux/libc_fatal.c:189
#4  0xb7e8af44 in *__GI___libc_fatal (message=0xb7fc75ec "libgcc_s.so.1 must be installed for pthread_cancel to work\n") at ../sysdeps/unix/sysv/linux/libc_fatal.c:200
#5  0xb7fc4ffa in pthread_cancel_init () at ../nptl/sysdeps/pthread/unwind-forcedunwind.c:65
#6  0xb7fc509d in _Unwind_ForcedUnwind (exc=0xb7b0edc0, stop=0xb7fc2bf0 <unwind_stop>, stop_argument=0xb7b0e454) at ../nptl/sysdeps/pthread/unwind-forcedunwind.c:126
#7  0xb7fc2b98 in *__GI___pthread_unwind (buf=<optimized out>) at unwind.c:130
#8  0xb7fbcce0 in __do_cancel () at pthreadP.h:265
#9  __pthread_exit (value=0x0) at pthread_exit.c:30
#10 0x08132ced in PyThread_exit_thread () at Python/thread_pthread.h:266
#11 0x08137c37 in t_bootstrap (boot_raw=0x8318aa8) at ./Modules/_threadmodule.c:1023
#12 0xb7fbbc39 in start_thread (arg=0xb7b0eb70) at pthread_create.c:304
#13 0xb7ef978e in clone () at ../sysdeps/unix/sysv/linux/i386/clone.S:130
"""

So if you're unlucky and end up closing the FD referring to libgcc used by dlopen/dlsym, you're pretty much screwed, and there's no way around this.

Note that you specific problem (upon PyThread_delete_key()) doesn't occur for python 2.7 because it uses and ad-hoc TLS implementation, whereas Python 3 uses the platform native TLS.
But as noted above, you can very well get an abort on pthread_exit(), and in likely many other places (pretty much everywhere libgcc can be dlopen'ed).

Unfortunately, we can't do much against this, so I'm tempted to close this as "wont fix".