Add ops-cpp target to torchvision (#3350)

Summary:
This diff adds a new target to torchvision which enables users to use torchvision ops from C++.

For now, the `cpp_library` is not used by the `python_cpp_library`. We should instead refactor the logic in torchvision to directly use `cpp_library` instead.

There is currently an inconsistency between fbcode and OSS users. OSS users can import torchvision via
```
#include <torchvision/vision.h>
```
while fbcode users need to do
```
#include <torchvision/csrc/vision.h>
```
It would be good to fix this discrepancy in the future.

I didn't directly use `test_frcnn_tracing.cpp` due to complications for getting the `.pt` file in a way that works for both OSS and fbcode, so instead we added a self-contained test that should validate that the torchvision ops are properly registered and visible to JIT

Reviewed By: datumbox

Differential Revision: D26225669

fbshipit-source-id: 5dd9fb98dd58e854f95806e4860d02f54fc04ea4

Co-authored-by: Francisco Massa <fmassa@fb.com>

test/cpp/test_custom_operators.cpp

@@ -0,0 +1,30 @@
+// Copyright 2004-present Facebook. All Rights Reserved.
+
+#include <gtest/gtest.h>
+#include <torch/script.h>
+#include <torch/torch.h>
+
+// FIXME: the include path differs from OSS due to the extra csrc
+#include <torchvision/csrc/ops/nms.h>
+
+TEST(test_custom_operators, nms) {
+  // make sure that the torchvision ops are visible to the jit interpreter
+  auto& ops = torch::jit::getAllOperatorsFor(torch::jit::Symbol::fromQualString("torchvision::nms"));
+  ASSERT_EQ(ops.size(), 1);
+
+  auto& op = ops.front();
+  ASSERT_EQ(op->schema().name(), "torchvision::nms");
+
+  torch::jit::Stack stack;
+  at::Tensor boxes = at::rand({50, 4}), scores = at::rand({50});
+  double thresh = 0.7;
+
+  torch::jit::push(stack, boxes, scores, thresh);
+  op->getOperation()(&stack);
+  at::Tensor output_jit;
+  torch::jit::pop(stack, output_jit);
+
+  at::Tensor output = vision::ops::nms(boxes, scores, thresh);
+  ASSERT_TRUE(output_jit.allclose(output));
+
+}