[XLA:GPU] Load kernel thunks' kernels before running them.

[XLA:GPU] Load kernel thunks' kernels before running them.

The motivation here is that with --xla_hlo_profile, we count the time
spent in Thunk::ExecuteOnStream, but we don't want to count the time
spent loading the CUDA code into the GPU as time spent in the first
kernel thunk we try to run.

PiperOrigin-RevId: 196314733