Use iterators in create_border_luma, add_residue, predict_dcpred

[okaneco]

Use iterators in places where indices are manually calculated because the compiler doesn't always optimize them. Iterators can remove extra bound checks or enable other optimzations like memset/memcpy or vectorized mov instructions.

Use branchless clamping in a loop to produce better vectorized code. This generates saturating truncation instructions instead of branches of greater than comparisons masking with and/andnot.

---

I was working on personal code and came across similar patterns where I found indexing manually to be a fair amount slower than using iterators.

I was reading the RFC to get a better understanding of the prediction functions and saw some places that could be improved.

The main heuristic I used for finding and replacing indexing was if it was complex enough (ie, more than for i in 0..arr.len() { arr[i] ... }), and the indexing was being done horizontally as opposed to vertical strides. Then I plugged the code in to Compiler Explorer to see if there were improvements, then testing it on the benchmark. I know that isn't the full picture because most of these functions have constants provided as arguments but it was still a helpful barometer.

There are 3 commits which get ~6% increase on the image benches. I've tried to break them up into small, easy to review commits.