[NNVM] Bug fix Prevent fusing convolution with injective op

[masahi]

Continuing from #1603.

Modify operator fuser to include following logics.

• During the initial partition step, if injective op is followed by broadcast op, mark broadcast op's op pattern to be injective
• In the last grouping step, check for input nodes op patterns. If one of them is kOutEWiseFusable and the other is kInjective, ignore the kOutEWiseFusable node and fuse only kInjective node.

A test case is included. This test fails without this PR.
I confirmed that this change does not affect compling resnet, mobilenet and vgg.

@tqchen please check if the logic is good.

[tqchen]

@merrymercy can you also review this?

[tqchen]

style, multiline if always enclose with {}

[tqchen]

Thanks @masahi , this is now merged

[kazum]

The below script fails after this commit. This kind of graph pattern can be found, for example, in NASNet.

import nnvm
from nnvm import symbol as sym

ch = 3
size = 8
data = sym.Variable(name="data")
concat = sym.concatenate(data, data)
conv = sym.conv2d(data=concat, kernel_size=(1,1), channels=ch*2)
net = sym.elemwise_add(concat, conv)
nnvm.compiler.build(net, "llvm", {"data": (1, ch, size, size)})

The result is as follows.

Traceback (most recent call last):
  File "tmp.py", line 10, in <module>
    nnvm.compiler.build(net, "llvm", {"data": (1, ch, size, size)})
  File "/home/kazutaka/git/tvm/nnvm/python/nnvm/compiler/build_module.py", line 304, in build
    graph = graph.apply("GraphCompile")
  File "/home/kazutaka/git/tvm/nnvm/python/nnvm/graph.py", line 234, in apply
    check_call(_LIB.NNGraphApplyPasses(self.handle, npass, cpass, ctypes.byref(ghandle)))
  File "/home/kazutaka/git/tvm/nnvm/python/nnvm/_base.py", line 75, in check_call
    raise NNVMError(py_str(_LIB.NNGetLastError()))
nnvm._base.NNVMError: [05:30:54] /home/kazutaka/git/tvm/nnvm/include/nnvm/././op.h:530: Check failed: op != nullptr

@masahi, can you check this out?

[masahi]

@kazum Thanks! I sent a PR to fix this #1622.