v0.8.0

Changes:

• Make sure GPUs are freed directly if the gpu process is killed (e.g. by VSCode)
• Remove the claim command, since nobody uses it anyway and it interferes with the above
• Set CUDA_VISIBLE_DEVICES with numeric IDs, anything else confuses older pytorch versions (torch.cuda.device_count() returns 0)
• Fix NVML struct versioning
• Print idle time in seconds rather than in minutes
• Decrease idle timeout to 1 min