Fix sort_test build on toolchains without native float16

[dcsid]

sort_test.cc fails to build where hwy::float16_t is emulated: std::iota with T{0} needs construction from int, and -kInf needs unary minus (GCC 9–12, Clang 11–15, armv7, ppc64le, LoongArch64, Windows LLVM). Fixes #3112.

• Replace the iota with ConvertScalarTo<T>, and x == kInf || x == -kInf with ScalarAbs(x) == kInf.
• Guard the f16 instantiation with HWY_HAVE_FLOAT16 (as in CallAllSortTraits): f16 GetLane does not compile on LSX (noted in sort_test fails to build on toolchains without native float16 (since 636d64d8) #3112).

Errors reproduce under -DHWY_HAVE_SCALAR_F16_TYPE=0 and compile cleanly with the fix; sort_test --gtest_filter='*SelectWithNaN*' passes on AArch64.

[jan-wassenberg]

Thanks :) FYI #3109 fixes the missing #if. Let's wait for that one to land, then merge.

I'm curious which compiler you're targeting? HWY_HAVE_SCALAR_F16_OPERATORS is not universal, but ought to be widely available. Perhaps we can get that to 1, or maybe we should extend the emulation of hwy::float16_t.

[dcsid]

Rebased onto #3109, now just the ConvertScalarTo/ScalarAbs constructs.

No particular compiler, just hit it on the CI matrix. FWIW the emulated float16_t gaps were construction from int (std::iota's T{0}) and unary minus (-kInf); covering those via HWY_HAVE_SCALAR_F16_OPERATORS or the emulation would drop the need for the guard.