Merge #4845

4845: Fix perf regression facet strings r=ManyTheFish a=dureuill

Benchmarks between v1.9 and v1.10 show a performance regression of about x2 (+3dB regression) for most indexing workloads (+44s for hackernews).

[Benchmark interpretation in the engine weekly meeting](https://www.notion.so/meilisearch/Engine-weekly-4d49560d374c4a87b4e3d126a261d4a0?pvs=4#98a709683276450295fcfe1f8ea5cef3).

- Initial investigation pointed to #4819 as the origin of the regression.
- Further investigation points towards the hypernormalization of each facet value in `extract_facet_string_docids`
- Most of the slowdown is in `normalize_facet_strings`, and precisely in `detection.language()`.

This PR improves the situation (-10s compared with `main` for hackernews, so only +34s regression compared with `v1.9`) by skipping normalization when it can be skipped.

I'm not sure how to fix the root cause though. Should we skip facet locale normalization for now? Cc `@ManyTheFish` 

---

Tentative resolution options:

1. remove locale normalization from facet. I'm not sure why this is required, I believe we weren't doing this before, so maybe we can stop doing that again.
2. don't do language detection when it can be helped: won't help with the regressions in benchmark, but maybe we can skip language detection when the locales contain only one language?
3. use a faster language detection library: `@Kerollmops` told me about https://github.com/quickwit-oss/whichlang which bolsters x10 to x100 throughput compared with whatlang. Should we consider replacing whatlang with whichlang? Now I understand whichlang supports fewer languages than whatlang, so I also suggest:
4. use whichlang when the list of locales is empty (autodetection), or when it only contains locales that whichlang can detect. If the list of locales contains locales that whichlang *cannot* detect, **then** use whatlang instead.

---

> [!CAUTION]
> this PR contains a commit that adds detailed spans, that were used to detect which part of `extract_facet_string_docids` was taking too much time. As this commit adds spans that are called too often and adds 7s overhead, it should be removed before landing.

Co-authored-by: Louis Dureuil <louis@meilisearch.com>
Co-authored-by: ManyTheFish <many@meilisearch.com>

milli/src/search/facet/search.rs

@@ -339,10 +339,18 @@ impl ValuesCollection {
 fn normalize_facet_string(facet_string: &str, locales: Option<&[Language]>) -> String {
     let options = NormalizerOption { lossy: true, ..Default::default() };
     let mut detection = StrDetection::new(facet_string, locales);
+
+    // Detect the language of the facet string only if several locales are explicitly provided.
+    let language = match locales {
+        Some(&[language]) => Some(language),
+        Some(multiple_locales) if multiple_locales.len() > 1 => detection.language(),
+        _ => None,
+    };
+
     let token = Token {
         lemma: std::borrow::Cow::Borrowed(facet_string),
         script: detection.script(),
-        language: detection.language(),
+        language,
         ..Default::default()
     };