recipe: scipy 1.18.0 + flet-libopenblas 0.3.33

[ndonkoHenri]

Adds scipy (flet-dev/flet#3391) and its BLAS/LAPACK backend flet-libopenblas.

Approach

• Fortran-free: scipy 1.18 builds with -D_without-fortran=true; the only casualty is the deprecated scipy.odr. This removes the cross-Fortran toolchain blocker entirely — scipy then needs only a BLAS/LAPACK backend.
• OpenBLAS everywhere: A new flet-libopenblas recipe builds OpenBLAS with NOFORTRAN=1 (BLAS + the f2c-converted LAPACK, no Fortran compiler) as a static lib, and scipy links it via -Dblas=openblas -Dlapack=openblas on both platforms. This is scipy's most-tested BLAS path (standard sgemm_ mangling), unifies iOS and Android, and sidesteps Apple Accelerate's macOS-centric gaps on iOS. scipy-openblas32 isn't an option (no iOS/Android wheels).

Recipes

• flet-libopenblas 0.3.33 — build.sh cross-compiles OpenBLAS per arch (ARMV8 / NEHALEM / ARMV7), static-only, and rewrites openblas.pc to be relocatable (and drops its baked-in -lgfortran/-lpthread). Statically linked into scipy, so it is a build-time host dep only (not in scipy's Requires-Dist).
• scipy 1.18.0 — meson-python recipe; host deps numpy 2.4.6, pybind11, flet-libopenblas, and flet-libcpp-shared (Android C++ runtime).

Platform notes

iOS builds with no patches. Android needs four small, platform-guarded fixes for bionic/NDK gaps (each a no-op on iOS):

• -Duse-pythran=false — pythran 0.18's ndarray.hpp has a ref-qualifier overload the newer NDK clang rejects (affected modules fall back to their Cython versions).
• android-bionic-clog-cpow.patch — bionic only declares the C99 clog/cpow from API 26; shim them for __ANDROID_API__ < 26 via functions present at API 24.
• android-ducc-no-affinity.patch — scipy.fft's vendored ducc pins threads with the glibc-only pthread_{get,set}affinity_np; fall back to its no-affinity path on Android (CPU pinning is only an optimization).
• android-x86_64-boost-longdouble.patch — Android x86_64 uses 128-bit IEEE-quad long double (LDBL_MANT_DIG==113), not x86's 80-bit; gate boost.math's x86 branch on ==64 so it falls through to boost's existing binary128 path.

Forge fix

src/forge/build.py: the site-packages pkg-config glob only matched */share/pkgconfig, missing numpy's numpy/_core/lib/pkgconfig/numpy.pc, so scipy 1.18's dependency('numpy') couldn't find it. Added */lib/pkgconfig and */_core/lib/pkgconfig — a general win for any dependency('numpy') consumer.

Validation

• Builds green across the full matrix: iOS (iphoneos.arm64, iphonesimulator.arm64, iphonesimulator.x86_64) and Android (arm64-v8a, x86_64, armeabi-v7a).
• CI: build + mobile tests green on Android and iOS for 3.12.
• On-device verified (iOS simulator + Android arm64 emulator): import scipy, scipy.linalg.solve/svd (OpenBLAS LAPACK), scipy.fft (ducc), and scipy.special.gamma(0.5)==√π all correct.