Improve multi instances (#8)

If multiple RIFE() calls use the same model on the same GPU, the heavy model weights are only loaded into VRAM once.

`gpu_thread` limit is respected globally across all instances, preventing VRAM spikes from simultaneous execution.