Make the custom NEON code for quantized GEMV more resilient to possible compiler bugs,

as in https://bugs.llvm.org/show_bug.cgi?id=34945

On the latest Android NDKs I don't observe any difference, but this being opensource, it will get compiled by different compilers, so it's probably useful to make it a bit foolproof as we've seen these issues often in practice.

In a nutshell:  don't use arrays of NEON variables, use separate plain NEON variables. Corollary: don't try to use small fixed-size loops to reduce redundancy, as even though the compiler does unrolled them, code structured in this way isn't compiled in the same way as code that's entirely statically unrolled with separate variables as opposed to array entries.

PiperOrigin-RevId: 234198355