simplify back propagation (remove unnecessary sequence traversal)

[tf318]

The original recursive back propagation algorithm performs breadth-first visiting of Values in the graph, simply to add them to a linear sequence. The sequence is then reversed and iterated over to invoke the _backward() function in each Value.

It is possible to invoke the _backward() function in a single pass during the breadth-first visiting of Values, without creating an interim linear sequence. This PR simplifies the algorithm accordingly.

[conscell]

@tf318 This implementation fails tests:
FAILED test_engine.py::test_sanity_check - assert 128.0 == 46.0
FAILED test_engine.py::test_more_ops - assert 15.166180758017475 < 1e-06