We are seeing some failures when comparing the results
for eigh with shape polymorphism and without.
Normally, shape polymorphism should not change the HLO
so a golden comparison is not necessarily bad, even though
for eigh we should check for correctness of the results
rather than identity.
We need to investigate this further but meanwhile turn
off these tests. The changes introduced recently for
shape polymorphism for eigh are not affecting the
code paths in absence of shape polymorphism. So it
is appropriate to just turn off the tests, and add
an error that shape polymorphism for eigh on
GPU is not ready.
PiperOrigin-RevId: 529388749
40aa4e1781