pl/python: fix ref leak on elog

Attached is a simple patch that fixes a minor problem in PL/Python: in
PLy_spi_execute_plan(), we invoke the input function for a Postgres data
type. If that function elog's (which is quite possible), PL/Python leaks
references to two temporary Python objects. The fix seems easy, albeit a
bit ugly: the reference count on one of the temporary objects can be
decremented before calling the function. A pointer that points into the
second object is passed to the input function itself, so we have to
PG_CATCH() the elog and decrement the refcount then.

Since the problem might actually occur in practice and the fix seems to
have little chance of inducing a regression, I'm thinking of applying
this to both HEAD and back branches. Barring any objections I'll do that
later today or early tomorrow.

-Neil