[gpu] Use batched syev for lax.linalg.eigh

[jaro-sevcik]

Let us use a batched syev for jax.lax.linalg.eigh to speed up the function
for non-trivial batch sizes.

Fixes #31368

Depends on openxla/xla#30810

[hawkinsp]

This LGTM. Do you also want to change the C++ code to select jacobi always when passed "default" as the algorithm?

[hawkinsp]

There are some test failures. PTAL? Most look like small tolerance bumps but e.g. the rank deficient one seems large.