[Bug] CoreML frontend fails on NNVM with opt_level=3 and target=CUDA

Development
#1

As per https://github.com/dmlc/tvm/pull/2476/files#r251020916

I tried to debug a bit.

  • Relay is good
  • NNVM + LLVM target is good
  • ResNet 50 is good
  • NNVM + CUDA fails
    • Disable AlterOpLayout: fails
    • Disable FoldScaleAxis: pass

Anyone has idea?