* arm deconv matmul use gemm

* reduce gemm armv7 register uses

* mha arm use softmax fp16

* fuse im2col and packb tile

* update f43 coeffs

* arm convolution winograd unified elempack

* disable bf16s test atm

* test gnu inline asm off

* use full date for pypi package

* split version date string only for dylib

* stb_image_write 1.16

* stb_image v2.28

* update emsdk 3.1.28

* enable stb arm neon

* update doc

Co-authored-by: ncnnnnn <67086033+ncnnnnn@users.noreply.github.com>

* update winograd f43 coeffs

* rvv tanh rework

* fix warnings

* rebuild qemu

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

* find cpu cache size from sysfs

* android l3

* make g_thread_affinity_mask singleton

* global mask

* simplify cast fp16 avx512 dispatch

* define sse4.1 macro on msvc avx+

* mha is now permute and reshape free
* gemm user defined tile mnk param

* improve gemm packb threading
* optimize tile size
* profile winograd condition
* handle threads changes

* enable VK_KHR_portability_subset extension if device support it

Co-authored-by: w1ndseeker <w1ndseeker@users.noreply.github.com>

* fix doc, sync x86 gemm fix

Co-authored-by: EdVince <EdVince@users.noreply.github.com>
Co-authored-by: nihuini <nihuini@tencent.com>

* cmake determine target 32bit and 64bit
* include opt source with non-runtime cpu
* check compiler support gnu style inline assembly

* drop old sgemm code

* fix convdw test

* fix avx512 gemm

* optimize prefer sgemm condition

* get cpu l2 cache size and resolve gemm tile size

* optimize constant tile K

* fix per-core l2 cache detection, better macos cpu cluster topology discovery

* gemm x86 multiply alpha beta in post gemm stage, enable one_blob_only
* relax mnk multiple restrictions
* make square tiles in each thread
* sanitize num_threads changes