Summary:
This adds the capability for caffe2 to call c10 operators and adds a dummy c10 sigmoid op as a proof of concept.

I used this test script to make sure it works:

    from caffe2.python import workspace, model_helper
    import numpy as np

    data1 = np.random.rand(16, 100).astype(np.float32)
    workspace.FeedBlob("data1", data1)
    m = model_helper.ModelHelper(name="my net")
    sigmoid1 = m.net.C10Sigmoid_DontUseThisOpYet("data1", "sigmoid1")
    sigmoid2 = m.net.Sigmoid("data1", "sigmoid2")

    workspace.RunNetOnce(m.param_init_net)
    workspace.CreateNet(m.net)
    data1 = np.random.rand(16, 100).astype(np.float32)
    workspace.FeedBlob("data1", data1)
    workspace.RunNet(m.name, 1)

    print(workspace.FetchBlob("data1"))
    print(workspace.FetchBlob("sigmoid1"))
    print(workspace.FetchBlob("sigmoid2"))

(and check that both sigmoid outputs are the same)
Pull Request resolved: pytorch#8809

Differential Revision: D8810207

fbshipit-source-id: 1135bb3ff87e2131bb4d52bba3177cc07bfb4b06