Always enable WorkerRun.use_gpu on versions requiring GPUs

WorkerRuns on templates were found to have use_gpu disabled after the migration in !2351 (merged), even though their WorkerVersion had gpu_usage set to Required. That was because on template processes, Process.use_gpu was always False, and the migration relied on Process.use_gpu to make its decision.

I updated the migration to always set WorkerRun.use_gpu when a version requires GPU usage. When a version supports GPU usage, it is only enabled on the WorkerRun if it was enabled in the process, which was the previous behavior when we had use_gpu on processes.

I already ran the queries manually in preprod, prod and demo to apply this change and get MLProd unstuck.