Update operators keep tensors on GPU between ops (where possible)

[karlhigley]

• [ ] Create a numpy/cupy dispatch mechanism (like pandas/cudf in NVT)
• [ ] Apply DLpack to pass GPU tensors from Python back-end to other models
• [ ] Update FilterCandidates
• [ ] Update SoftmaxSampling
• [ ] Update Faiss and Feast ops to convert to GPU?

[karlhigley]

Depending on how this turns out, we may or may not find it worthwhile to add a graph optimizer to condense multiple operators into a single TritonPythonModel. It would still help us avoid the scheduling overhead associated with passing requests between models, but it might not be a big boost if it combining operators no longer helps us avoid GPU-CPU roundtrip conversions.