Depending on the arch, ldr/str can be faster than ld1/st1, especially
for loading to one lane form. For example, on Cortex A75,

1. execution latency of 'ldr q0' and 'ldr h0' are 5
2. execution latency of 'ld1 {v0.16b}' is 6
3. execution latency of 'ld1 {v0.h}[0]' is 8

On Cortex X3,
1. execution latency of 'ldr q0' and 'ldr h0' are 6
2. execution latency of 'ld1 {v0.16b}' is 6
3. execution latency of 'ld1 {v0.h}[0]' is 8

Signed-off-by: Zhao Zhili <zhilizhao@tencent.com>

* Fix tanh typo for rvv.

* Fix tanh for rvv fp16.

* fix potential fp16s bf16s conflicts on armv7 vfpv4

* but prefer fp16 on armv8.2

* benchmark: Don't use std::thread when NCNN_THREADS is OFF

* Add wasi support

cmake -B build -DCMAKE_BUILD_TYPE=Release \
	--toolchain ${wasi_sdk}/share/cmake/wasi-sdk.cmake \
	-DNCNN_RUNTIME_CPU=OFF \
	-DNCNN_DISABLE_EXCEPTION=ON \
	-DNCNN_THREADS=OFF

cmake --build build

After build, you can run benchncnn on cmdline with wasmtime:
wasmtime --dir . benchncnn

* update swiftshader

* skip vs2017 swiftshader

* test mha oom

The updated ruapu adds support for multiple architectures such as RISC-V, MIPS, and Loongson, and can detect more Arm features.

The latest version is 10b02b3755.

* skip mips loongarch riscv oom test atm

* test softmax oom

* workaround vs2019 crash

* Remove two potential warnings for VisualStudio

* fix mingw64 avx crash

* fix build issue in termux

When performing destroy_gpu_instance(), g_instance.created is always 0.

* windows clang ci

* clang msvc use x86intrin.h for xop

* test arm64 compiler features

Here is a big difference between C and C++. foo() in C means that
the funtion takes an unspecified number of arguments, while
foo(void) means the function takes no argument.

Fix -Wstrict-prototypes warning.

Signed-off-by: Zhao Zhili <zhilizhao@tencent.com>

* destroy_gpu_instance() will internally ensure that all vulkan devices are idle before proceeding with destruction.

* fix moltenvk static linking

* pnnx fix some undefined dtype

* fix ncnn convdw1d dynamic weight loading

with `/openmp:llvm` compile option

remove _sum0=_sum0

* Fix `destroy_gpu_instance` crash

* Additional check and clear

* promote vfpv4 for auto fp16 storage conversion

* always report neon and vfpv4 for arm64

* replace storexxx to vsseg2e32_v_f32m1

* refine transpose

---------

Co-authored-by: Xinyu302 <Xinyu302@users.noreply.github.com>