Perf: AArch64: U256::const_eq auto vectorization produces longer dependency chain than scalar lowering

For small fixed size equality checks (LIMBS = 4), LLVM currently vectorizes the reduction into a NEON horizontal reduction sequence:
```
ldp q0, q1, [x0]
ldp q2, q3, [x1]
cmeq v1.2d, v1.2d, v3.2d
cmeq v0.2d, v0.2d, v2.2d
uzp1 v0.4s, v0.4s, v1.4s
mvn v0.16b, v0.16b
umaxv s0, v0.4s
fmov w9, s0
bic w0, w8, w9
```
This creates a dependency chain(as it should) with long dependency chain

Now if we were to turn off auto vectorization we get this:
```
const_eq:
ldp x8, x11, [x1]
ldp x9, x10, [x0]
cmp x9, x8
ccmp x10, x11, #0, eq
ldp x8, x10, [x0, #16]
ldp x9, x11, [x1, #16]
ccmp x8, x9, #0, eq
ccmp x10, x11, #0, eq
cset w0, eq
ret
```
This creates a dependency chain(as it should) too but with way more exploitation of the ILP

Will you handroll it for aarch64 or maybe use an attribute(can't think of any)?

anyway the cycle difference between this is like 60+ but of course in a prod this won't be a standalone execution so it kinda frees up the pipeline faster and prevents CPU stalling on instructions