Handle Python layer exceptions correctly

[longjon]

(This is a simplified alternative implementation of part of the first commit of #2001, plus a test.)

As noted by @tnarihi, Python layer exceptions are not handled correctly in pycaffe. PyErr_Print is used to display exception text on stderr, but this has the side effect of clearing the exception, which results in an additional SystemError: NULL return without exception set.

The exceptions get passed through correctly if we simply don't try to catch and print them; that's the first commit. (Note that the Python interpreter will print these exceptions, if uncaught; it's only the caffe tool that won't.) Since we'd still like to see the exception text when using the caffe tool, the second commit adds a single, ifdefed, catch/PyErr_Print block to the tool code. (There's no harm in calling PyErr_Print since we know we're not in the interpreter.)

I prefer this approach to the one in #2001, for these reasons:

• it simplifies the Python wrapper code, removing the duplicate error catching blocks
• it removes the need to think about where exceptions might occur in future wrapper code
• it avoids the business of embedding Python calls to read (and restore) the exception
• calling code (from C++) has the option of avoiding the exception printout (e.g., and dealing with the exception)

The disadvantage of this approach is that:

• Python exceptions won't be automatically printed when invoking nets with Python layers from C++ code (outside the Python interpreter) other than the caffe tool

Let me know what you think @tnarihi and others, and thanks @tnarihi for pointing out this issue. (Note also that this does not include all of the exception-fixing functionality from #2001, which I'll plan to check out separately.)

[shelhamer]

Merging as I agree that this is simpler than the solution in #2001. Thanks Jon and thanks Takuya for raising the issue!