* pnnx convert onnx sdap

* test reduce

* pnnx convert onnx multiheadattention

* onnx reducemean reducesum

* reducemax reducemin reduceprod

* mask buggy torch

* avoid shadow output

* pnnx convert onnx layernorm

* fuse early

* skip layernorm affine false test for torch 2.1

* pnnx convert onnx layernorm instancenorm groupnorm

* take num_features from input shape for instancenorm module

* torch < 1.10 can not handle track_running_stats=True

* pnnx convert onnx avgpool

* fuse gather indices

* generalized conv convtranspose

* skip maxpool dilation test for torch < 1.12

* ceil mode for opset 9

* pnnx reset onnx input shape

* eliminate noop cast

* pnnx convert onnx pad

* pnnx convert onnx linear sigmoid softmax relu

* old onnx softmax

* fix ncnn tools crash with compound layer creation

* write gemm weight data

* pnnx fix some undefined dtype

* fix ncnn convdw1d dynamic weight loading

* split load torchscript sources from pnnx, drop cxxabi hack, link torchscript with whole-archive

* check compiler cxx11 abi

* eliminate some onnx noop

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* pnnx handle two operands add/sub/rsub variant

* fuse dynamic slice indexes, wip

* pnnx sliceindexes

* reset device may change non-dtype input numeric 5 to 6

* print inf as float

* preserve dtype for generation op

* pnnx convert torch.masked_select

* test masked_select

* test negative slice

Signed-off-by: hugo-syn <hugo.vincent@synacktiv.com>

* pnnx handle more softmin logsoftmax dtype, fuse static full range slices to tensor_split

* fix convert nn.Conv2d with none bias tensor

* fix embedding input with batch index zero