[XLA:GPU] Don't run convolution autotuning concurrently on the same GPU.

This lets us compile two independent HLO modules concurrently without
them stepping on each others' toes.

That's *all* it lets us do; the lock doesn't prevent other processes on
the machine or even other things within the same process from using the
GPU concurrently.

PiperOrigin-RevId: 207624100