If we have a chain of generators/coroutines that are 'yield from'ing each other, then resuming the stack works like:

- call send() on the outermost generator
- this enters _PyEval_EvalFrameDefault, which re-executes the YIELD_FROM opcode
- which calls send() on the next generator
- which enters _PyEval_EvalFrameDefault, which re-executes the YIELD_FROM opcode
- ...etc.

However, every time we enter _PyEval_EvalFrameDefault, the first thing we do is to check for pending signals, and if there are any then we run the signal handler. And if it raises an exception, then we immediately propagate that exception *instead* of starting to execute bytecode. This means that e.g. a SIGINT at the wrong moment can "break the chain" – it can be raised in the middle of our yield from chain, with the bottom part of the stack abandoned for the garbage collector.

The fix is pretty simple: there's already a special case in _PyEval_EvalFrameEx where it skips running signal handlers if the next opcode is SETUP_FINALLY. (I don't see how this accomplishes anything useful, but that's another story.) If we extend this check to also skip running signal handlers when the next opcode is YIELD_FROM, then that closes the hole – now the exception can only be raised at the innermost stack frame.

This shouldn't have any performance implications, because the opcode check happens inside the "slow path" after we've already determined that there's a pending signal or something similar for us to process; the vast majority of the time this isn't true.

I'll post a PR in a few minutes that has a test case that demonstrates the problem and fails on current master, plus the fix.