when using `/arch=AVX` or similar then SSE4.2 does not get activated (on MSVC at least, but should be similar on g++/clang)

This PR fixes this.

Fixes errors like

FAILED: test/x86/avx512/cast-emul-c (also load, loadu, max, min, permutexvar, reduce, set, setzero, set1, storeu, f16c)
clang-17  -o test/x86/avx512/cast-emul-c test/x86/avx512/cast-emul-c.p/cast.c.o -Wl,--as-needed -Wl,--no-undefined --target=riscv64-linux-gnu -Wl,--start-group -lm -Wl,--end-group
/usr/bin/riscv64-linux-gnu-ld: test/x86/avx512/cast-emul-c.p/cast.c.o: in function `simde_test_x86_assert_equal_f16x32_':
/opt/simde/riscv64-clang-17/../../../usr/local/src/simde/test/x86/avx512/test-avx512.h:14: undefined reference to `__truncsfhf2'

FAILED: test/x86/avx512/cmp-emul-c (also reduce)
clang-17  -o test/x86/avx512/cmp-emul-c test/x86/avx512/cmp-emul-c.p/cmp.c.o -Wl,--as-needed -Wl,--no-undefined --target=riscv64-linux-gnu -Wl,--start-group -lm -Wl,--end-group
/usr/bin/riscv64-linux-gnu-ld: test/x86/avx512/cmp-emul-c.p/cmp.c.o: in function `simde_mm512_cmp_ph_mask':
/opt/simde/riscv64-clang-17/../../../usr/local/src/simde/simde/x86/avx512/cmp.h:741: undefined reference to `__extendhfsf2'
/usr/bin/riscv64-linux-gnu-ld: /opt/simde/riscv64-clang-17/../../../usr/local/src/simde/simde/x86/avx512/cmp.h:741: undefined reference to `__extendhfsf2'
/usr/bin/riscv64-linux-gnu-ld: /opt/simde/riscv64-clang-17/../../../usr/local/src/simde/simde/x86/avx512/cmp.h:741: undefined reference to `__extendhfsf2'
/usr/bin/riscv64-linux-gnu-ld: /opt/simde/riscv64-clang-17/../../../usr/local/src/simde/simde/x86/avx512/cmp.h:741: undefined reference to `__extendhfsf2'
/usr/bin/riscv64-linux-gnu-ld: /opt/simde/riscv64-clang-17/../../../usr/local/src/simde/simde/x86/avx512/cmp.h:741: undefined reference to `__extendhfsf2'
/usr/bin/riscv64-linux-gnu-ld: test/x86/avx512/cmp-emul-c.p/cmp.c.o:/opt/simde/riscv64-clang-17/../../../usr/local/src/simde/simde/x86/avx512/cmp.h:741: more undefined references to `__extendhfsf2' follow
clang: error: linker command failed with exit code 1 (use -v to see invocation)

sse/sse2: added ".f16" float16 accessor for m128 & m128i

mm{256,512}_cvt{ph_ps,ps_ph}: use .f16 instead of converting from u16

mm512_cvt{epi8_epi16,epi32_ps,cvtph_ps}: added _mm256_cvt* native fallbacks

https://github.com/llvm/llvm-project/commit/f253bb640d97756d2808bb7c7b2bb31b1090a654
https://reviews.llvm.org/D144169

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105339

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98521

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94482

Clang incorrectly sets the alignment bit on 32-bit for the multiple load
instructions. This results in a bus error on hardware.

Similar to _mm_shuffle_epi8, qtbl/qtbx now split up to multiple shuffles on
A32V7. With multi-vector shuffles, two tables are created, the indexes
are subtracted, and either TBX'd on the result of the other table or ORR'd
together (TBX is mandatory for qtbx, but ORR is chosen when the lookups would
not be theoretically dual issued).

Caught by clang-14+

_make
_const
_extract_lane
_splat
_const_splat (also signed variants)

My attempt for the unsigned versions did not work

Fixes: #1008