Shared memory option

[cooperlab]

Provide a shared memory option when creating an instance of the Requests class.

This would allow Requests to make GRPC requests using shared memory, as shown in the example here: https://github.com/triton-inference-server/client/blob/main/src/python/examples/simple_grpc_shm_client.py.

With non-uniform batch sizes we need to allocate new shm resources for each request. Otherwise, these can be reused. We can provide a flag to indicate whether batch size is non-uniform.

[cooperlab]

See triton-inference-server/server#3929 for discussion of shared memory re-use. OP suggests no observable overhead from creating a new shm for every request. This would make supporting non-uniform batch size easy.

[create-issue-branch]

Branch issue-24-Shared_memory_option created!