Begin defaulting GPUOptions.experimental.pending_cap to 8. This

has the effect of limiting the maximum number of pending GPU kernels
(when using the default single compute stream configuration) to
no more than 8.  Testing with benchmark programs in single-GPU and
distributed configurations suggests this always yields better performance
than unlimited queuing, probabaly because of effects on GPU/CPU i/o
which needs to wait on pending kernels for memory allocation safety.

PiperOrigin-RevId: 233456126