improve reflect replicated pad pattern

* added erf implementation

* added testcase for erf

* added onnx2ncnn support of erf

* reset device check dtype kind int

* placeholder for ncnn sign

* convert torch maximum minimum

* torch.max as expression

* torch.min as expression

* Add logxxx to log comp xxx rewriter

* Use pattern matching for LogSigmoid and LogSoftmax

* Add conversion passes for functional counterparts

* Update documentation

Tensor.fill
Tensor.index_put
Tensor.to
Tensor.type_as
torch.topk
fmod
call Tensor member functions with inputnames
static shape_as_tensor
nn.Linear dynamic bias
eliminate noop type_as
convert two-dim nn.Linear to ncnn gemm
convert torch.stack to ncnn concat+reshape
ignore torch einsum path input

* fix build, prepend batch for broadcast reshape

* sanitize filename

* do not fuse to eltwise if broadcast

* update riscv qemu

* c906 test on qemu

* fix qemu aarch64

* simplify binaryop

* less gpu test

* update binaryop broadcast doc

* do not test atan2 zero

* write shape, fuse sam image encoder attention

* set more dynamic shape as static

* less warning for constant tensor node

fuse F.scaled_dot_product_attention

* pnnx fuse transformer clip attention

* skip fuse mha for 1.8

* select one method other than forward

* pnnx fuse diffusers attentionblock

* pnnx convert sdpa

* pnnx fuse diffuser attention2