[XLA:GPU] Pass all four args to custom-call convs when they're created.

[XLA:GPU] Pass all four args to custom-call convs when they're created.

A custom-call-conv should have four arguments: lhs, rhs, algorithm, and
use-tensor-cores.  CudnnAlgorithmPicker did the right thing, and that
path is exercised 99% of the time.  But CudnnAlgorithmPicker can fail,
and if it does, we're stuck with whatever we had before.  What we had
before only had three of the four args, which is bad.

In addition to fixing it, added an e2e test that catches the bug.

PiperOrigin-RevId: 192694119