use combine_x for sse/avx vector combination

[nihui]

No description provided.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 94.75655% with 14 lines in your changes missing coverage. Please review.

Project coverage is 95.59%. Comparing base (23b64c9) to head (0e54b5c).
Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layer/x86/convolution_3x3_winograd.h	86.04%	12 Missing ⚠️
src/layer/x86/innerproduct_fp.h	87.50%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6113      +/-   ##
==========================================
- Coverage   95.59%   95.59%   -0.01%     
==========================================
  Files         829      829              
  Lines      270462   270452      -10     
==========================================
- Hits       258556   258546      -10     
  Misses      11906    11906              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

The binary size change of libncnn.so (bytes)

architecture	base size	pr size	difference
x86_64	16495480	16495480	0 😘
armhf	7384096	7384096	0 😘
aarch64	10777496	10777496	0 😘

[Copilot]

Pull Request Overview

This PR consolidates manual AVX/SSE insert intrinsics into unified combine* helper functions for cleaner vector concatenation.

• Introduced #if __AVX2__ fallback in combine4x2_epi32 and added new combine helpers.
• Replaced direct _mm*_insert* calls with combine4x2_ps, combine4x2_epi32, combine8x2_ps, and combine4x4_ps across x86 layer implementations.
• Added #include "x86_usability.h" where helper functions are used.

Reviewed Changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/layer/x86/x86_usability.h	Added AVX2 guard to combine4x2_epi32 with fallback.
src/layer/x86/unaryop_x86.cpp	Replaced _mm512_insertf32x8 with combine8x2_ps.
src/layer/x86/shufflechannel_x86.cpp	Added include and used combine4x2_ps.
src/layer/x86/scale_x86.cpp	Replaced multiple insert intrinsics with combine helpers.
src/layer/x86/rmsnorm_x86.cpp	Swapped insert calls for combine4x2_ps/combine8x2_ps.
src/layer/x86/prelu_x86.cpp	Used combine helpers in slope loading.
src/layer/x86/lstm_int8.h	Replaced integer insert with combine4x2_epi32.
src/layer/x86/layernorm_x86.cpp	Replaced insert calls with combine helpers.
src/layer/x86/innerproduct_gemm_fp.h	Replaced _mm256_insertf128_ps/_mm512_insertf32x8 with combine.
src/layer/x86/innerproduct_fp.h	Swapped out insert intrinsics for combine calls.
src/layer/x86/groupnorm_x86.cpp	Used combine helpers for mean/var broadcasts.
src/layer/x86/gemm_x86.cpp	Replaced insert intrinsics in transpose-pack with combine.
src/layer/x86/convolution_packed_int8.h	Replaced integer insert calls with combine4x2_epi32.
src/layer/x86/convolution_packed.h	Used combine* helpers instead of nested insert intrinsics.
src/layer/x86/convolution_3x3_winograd_int8.h	Swapped _mm512_inserti32x8 with combine8x2_epi32.
src/layer/x86/convolution_3x3_winograd.h	Replaced insert intrinsics in Winograd transforms.
src/layer/x86/convolution1d_packed.h	Updated insert intrinsics to combine helpers.
src/layer/x86/binaryop_x86.cpp	Replaced broadcast insert calls with combine helpers.
src/layer/x86/batchnorm_x86.cpp	Used combine* helpers for parameter loading.

Comments suppressed due to low confidence (1)

src/layer/x86/x86_usability.h:888

• The fallback branch for combine4x2_epi32 uses _mm256_insertf128_si256, which appears to be the float-version intrinsic. It should use the integer insert intrinsic _mm256_inserti128_si256 to correctly handle __m256i vectors.

return _mm256_insertf128_si256(_mm256_castsi128_si256(a), b, 1);