`initialize_kernels` is prone to becoming wildly contended

[slomp]

I have observed that halide_*_initialize_kernels() always acquires a lock, even if the object already exists. When multiple threads call into the runtime concurrently, initialize_kernels() keeps getting called, even when the void** state_ptr has already been initialized. This turns out to be a fairly obtuse contention point, which can be eliminated by atomically loading state_ptr first and checking if it is null (and thus actually require initialization.