[release/3.4][Build] Packaging vendored liblapack dylib on ARM64 platform

[risemeup1111]

PR Category

Environment Adaptation

PR Types

Bug fixes

Description

ARM64 上编译单独编译 liblapack 动态库，避免 paddle arm64 wheel 包里 liblapack 及其依赖项均为 x86 的

dylibs in paddle/libs	arch
libblas.so.3	x86
libgfortran.so.3	x86
libphi_core.so	arm64
libphi.so	arm64
libwarpctc.so	arm64
libcommon.so	arm64
liblapack.so.3	x86
libphi_gpu.so	arm64
libquadmath.so.0	x86
libwarprnnt.so	arm64

该问题如下代码即可复现：

python -c 'import paddle; x = paddle.to_tensor([[1, -2j], [2j, 5]]).cpu(); out_value = paddle.eigvalsh(x, UPLO="L"); print(out_value)'

是否引起精度变化

否

^{This PR is co-authored with @codex (gpt-5.6 sol max)}

---

Cherry-pick of #79516 (authored by @SigureMo) to release/3.4.

devPR:#79516

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[risemeup1]

LGTM