[XLA:GPU] Rewrite elemental emission of bitcasts

My first attempt at this only handled bitcasts that implement a reshape
operation, now transposes or mixed bitcasts are handled as well. There is
probably some optimization potential to reduce the amount of address arithmetic
emitted to IR for a follow-up.

This is already tested fairly well with the existing test suite, there are
failing tests with layout_assignment before fusion without this change.

PiperOrigin-RevId: 188155082