Fix stale output type bugs in rewrites

[ricardoV94]

Follow up to #2169 (fixed by #2167).

I had claude do an exhaustive search of rewrites that may do similar fatal/wrong reasoning based on output.type that may become stale. Narrowed down to rewrites where it would return invalid results or raise errors.

It found two, which are fixed here with regression tests.

This is a symptom of a bad design. Broadcasting in always implicit from the input static types. If an input has static shape 1 along a dimension it is allowed to broadcast with other inputs by the Op. If it has None it is not. Rewrites though may replace a None dimension by a 1, and a graph that did not allow an input to broadcast now allows (see #408 for a suggested solution).

For rewrites, it means we can't do either of the following:

1. Assume recreating the same node or equivalent one will have the same static output broadcastability.
2. Look at the delta between output broadcastable and input broadcastable to gauge whether broadcasting is happening. Input shape of 1, output shape of None doesn't mean there is another None input that broadcasts it, it may be that all inputs are now 1 but the node is stale.

[ricardoV94]

Agent report