Automatic update of fbcode/onnx to f461f7aad9987635b4aff108620ed7918f002d19

[houseroad]

Summary:
Previous import was 882c5283c54345d131e8fe5c859e4844dcf7ca8e

Included changes:

• f461f7a: Show the op's type and name when the shape inference is failed. (incompatibility between Pytorch 0.1.12 and cuda 8.0.61/cudnn 5.1.10 ? #1623)
• ab8aaf9: Add scan test case (Check cuDNN version at runtime #1586) <G. Ramalingam>
• c95357e: link the tutorial (merge from origin #1650)
• d7e2420: Upgrade label encoder to support more input types (Rewrite documentation on torch.autograd.backward() #1596)
• 6425108: Add Doc about Adding New Operator into ONNX (Don't visit grad_fn in THPVariable_traverse #1647)
• 295889c: use an empty initializer to create map (Conv double backward #1643)
• e38f3ec: Remove redundant const (python setup.py install get error #1639)
• ea694bf: implement fuse reduce->unsqueeze + fix assumption in nop_dropout pass (Implement transposed dilated convolutions #1565)
• 6db386e: make output shape clear enough for Softmax family (Variable.clone() does not clone to the same device #1634)
• 2b67c6e: fix batchnorm doc (Error during import torch, NameError: name '_C' is not defined #1633)
• c901784: remove inappropriate consts (Add scatterAdd and fix scatter grad formula #1632)
• de82119: Shape inference fix for broadcast, concat and scan (cudnnFindEx #1594)
• d7ffe3b: Update Optimizer Docs (Conv2d is missing CUDA/CPU type checks #1607)
• d09d139: mark PROTOBUF_INCLUDE_DIRS as BUILD_INTERFACE (fix corner-case in MaxPooling #1466)
• eb4b7c2: allow variadic parameters of different types (Have torch.manual_seed seed all GPUs as well #1615) <G. Ramalingam>
• 4166246: Fix onnxifi test (Fix indices bug in MM.py (#1613) #1617)
• 6706a4d: Fix a bug in vector address access (more explicit representation abstract class #1598)
• ae39866: Separate types of inputs 1 and 2 in OneHot op. (Fix a bug when transferring tensors from other devices to device 0 #1610)
• 45ba661: Handle new types in the switch. (torch.cuda.device not working but torch.cuda.set_device works #1608)
• 14853b6: Bump docker image version to 230 used in CircleCI (giving a 3d input to Conv2d gives a bad error message #1606)
• e0993b8: [onnxifi] Make sure that backend handles run async. (Add argmax and argmin to docs #1599)
• e6965cc: Introduce SparseTensor ML proto (Variable-size input #1554)
• 75b782f: In driver test check the return status of onnxGetBackendIDs (unexpected keyword argument 'keepdim' #1597)
• c05b364: Make CI log less verbose (DataLoader workers deadlocked #1595)
• fa568e4: Loop type shape inferencing (Pad PackedSequences to original batch length #1591)
• 937e64c: add uint8 (Fix Tensor.set_ with kwargs #1590)
• f86e951: Add domain as an optional parameter for make_node function (Advanced Indexing Part 1 -- Purely Integer Array Indexing #1588)
• ff45588: Remove unreachable code in shape_inference.h (Fix error message in CUDA forked subprocess #1585)
• f7dcad0: Add several hyperbolic function ops. (DataParallel does not effectively use GPU RAM on device_id 1 #1499)
• a60ac7d: Add OneHot op to ONNX. ([Feature Request] Asynchronous torch.save() #1567)
• f6c3a7e: [compiler flag] Issue a warning if class has virtual method but missing virtual dtor. (Add Conv2dLocal module #1583)
• 88d1784: Fix MaxUnpool shape inference when output_shape is provided as input (improve the argument types and return types in nn docs #1578)
• 20041b7: Add type shape inferencing for the If operator (Gradient size mismatch for nn.Bilinear #1571)
• d6c4c75: Add a virtual destructor to GraphInferencer (.gitignore updated with editor temporaries #1574)
• a339598: fix ConvTranspose spec (better _calculate_fan_in_and_fan_out #1566)

Differential Revision: D13263831

[zrphercule]

LGTM