perf: improve partial_lookahead

[ritchie46]

Use the previous partition point as educated guess for the first partition in binary search. I assume this fixes #11243

Still have to benchmark.

[ritchie46]

@MarcoGorelli can you take a look if this improves your cases? I don't see much of a difference. 🤔

[MarcoGorelli]

From having tried this out on a kaggle notebook it looks like it's got slightly slower https://www.kaggle.com/code/marcogorelli/polars-timing?scriptVersionId=144087899

But the difference between the two function calls doesn't seem very large there to begin with, perhaps because they only have 2 cores there, will check later

[orlp]

I don't think this really helps, and we should be doing this in a single linear scan anyway. In pseudo-code:

let mut i = start_offset;
let mut j = start_offset;

let mut results = Vec::with_capacity(upper_bound - lower_bound);
while i < upper_bound {
    if j < upper_bound && time[j] < time[i] + period {
        j += 1;
    } else {
        results.push((i, j-i));
        i += 1;
    }
}