glsl has no string type anyway

* NCNN_LOGE in glsl

* Update glsl-extension.md

* check image fp16

* clean vulkan shader common extension

* macro suffix makes glslang unhappy

* cmake check compiler test cannot be optimized out
* drop requant pack4

* restrict one dim quantize scale size

* sse2 requantize pack8

* remove dequantize pack8 test, seems to be useless

* check compiler supports isa with optimization enabled

* skip round problem
* sde on ubuntu24

* zfh zvfh xtheadvector infra
* dispatch for rvv and xtheadvector
* dispatch for non-vector zfh
* port xtheadvector recp rsqrt trunc
* general rvv gemm
* c906 and c910 ci
* old tuple code clean
* update riscv64 ci
* update build doc
* drop old th1520 toolchain

* only supports hann, hamming and all-one window
* inverse spectrogram does not support length parameter
* spectrogram always returns torch.view_as_real(out) as ncnn does not support complex typed mat yet
* inverse spectrogram always accepts torch.view_as_complex(in) as ncnn does not support complex typed mat yet

ref https://gitlab.freedesktop.org/mesa/mesa/-/issues/10847

* x86 vulkan fallback
* comment about bf16s

* quantize gemm

* write gemm quantize scales

* update doc

* less openmp args

* x86 riscv fallback

* skip gemm vulkan int8

* fix noint8 test, fix arm bf16 test

* enable vfpv4 on neon build only

* fix gemm vulkan without C

* fp16 pack8 output

* enable elempack=8 only for asimdhp+

* tiled gemm int8 test

* opt arm64 tiles, fix asimdhp dispatch

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>