Memory corruption

[eholk]

This is non-deterministic, but we're starting to see it in a couple of test cases and programs now (e.g. #147).

One that shows it pretty reliably is test/read-file.kfc. Here is a stack trace for one of the failures:

* thread #1: tid = 0x1451166, 0x00007fff8ba3aa90 libc++abi.dylib`__dynamic_cast + 33, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0xfffffffffffffff0)
  * frame #0: 0x00007fff8ba3aa90 libc++abi.dylib`__dynamic_cast + 33
    frame #1: 0x00007fff91ba8e38 OpenCL`___lldb_unnamed_function1037$$OpenCL + 96
    frame #2: 0x00007fff91ba8e96 OpenCL`___lldb_unnamed_function1039$$OpenCL + 14
    frame #3: 0x00007fff91ba9264 OpenCL`___lldb_unnamed_function1050$$OpenCL + 62
    frame #4: 0x00007fff91ba92b0 OpenCL`___lldb_unnamed_function1052$$OpenCL + 14
    frame #5: 0x00007fff91b8e5c5 OpenCL`clReleaseProgram + 23
    frame #6: 0x00007fff8305f8cd libsystem_c.dylib`__cxa_finalize_ranges + 322
    frame #7: 0x00007fff8305fbd0 libsystem_c.dylib`exit + 55
    frame #8: 0x00007fff869425d0 libdyld.dylib`start + 8
    frame #9: 0x00007fff869425c9 libdyld.dylib`start + 1

I would guess we have a race condition in how OpenCL resources are shutting down. Maybe we aren't reference counting correctly somewhere.