[XLA:GPU] Fuse scalar constants

This doesn't change codegen directly, but makes dealing with scalar broadcasts
much easier and the graph easier to read. This required changing the dot *
alpha fusion logic quite a bit, but I think for the better.

The emitter change is a bit of a hack. The more I look at this code the more
broken it seems. Need to find a more sustainable way of emitting what is
essentially a memset.

PiperOrigin-RevId: 200111599