Use load_array_ref to implement from_slice

[valadaptive]

I meant to use this method to implement the from_slice method, but apparently used load_array instead.

[LaurenzV]

Perhaps this will fix our issues with #171!

[Shnatsel]

This fixed much of my issues with QuState/PhastFT#58 !

There's still a performance gap vs wide but this closes much of it!

[LaurenzV]

How much is left?

[Shnatsel]

fearless_simd is 7% to 13% slower on Apple M4 depending on the benchmark (based on a quick run, not a full run; I can do a full run with more tests later).

On x86 (Zen4) it ranges from on par to 6% worse but that's not perfectly apples-to-apples since Zen4 has AVX-512 (emulated, double-pumped) that wide can use but fearless_simd cannot so I'm not too worried about that.