fix GPU job agent requirements in pipeline

[juliasloan25]

In our buildkite longruns pipeline, we specify agent requirements for the GPU AMIP runs (introduced in #632) that are very difficult (impossible?) to satisfy. Because of this, these jobs are waiting for an agent for a very long time (see here). We need to update these jobs to request GPUs and an appropriate amount of memory, but not multiple nodes/tasks. We should also figure out what the appropriate amount of memory is.

Update: the jobs take < 5GB of memory when run for a few timesteps, so the existing limits of 16-20GB should be sufficient (checked using CUDA.memory_status).