[GPU] Extend reflection to retrieve vector values

[ldematte]

While running gpu:yamlRestTest tests, I realized that under some circumstances the scorer is not Lucene99MemorySegmentFloatVectorScorerSupplier, but DefaultFlatVectorScorer$FloatScoringSupplier. This PR extend reflection to support the "fast path" to access vector values to this class too.

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[mayya-sharipova]

@ldematte How do encounter DefaultFlatVectorScorer$FloatScoringSupplier?

I checked Lucene code, and DefaultFlatVectorScorer$FloatScoringSupplier does NOT implement HasIndexSlice, so we will always return null and fall back to generating merged float values directly:

FloatVectorValues floatVectorValues = MergedVectorValues.mergeFloatVectorValues(fieldInfo, mergeState);

So I think the changes in this PR are NOT necessary.

---

What do you think?

[ldematte]

I checked Lucene code, and DefaultFlatVectorScorer$FloatScoringSupplier does NOT implement HasIndexSlice, so we will always return null

It's not the supplier that needs to implement HasIndexSlice, but the vectorValues it returns. As you can see in the picture, DefaultFlatVectorScorer$FloatScoringSupplier is returning a OffHeapFloatVectorValues, which implements slices/memorysegments.

You do see DefaultFlatVectorScorer$FloatScoringSupplier in YAML rest test; while rest tests do use a "real" cluster, there are differences, so I'm not sure how/if we can see the same in a full distribution. Still, it may happen, so I think it's worth adding.

[mayya-sharipova]

@ldematte Thanks for further clarifications.

Sorry I was checking how Lucene99MemorySegmentFlatVectorsScorer falls back to DefaultFlatVectorScorer::getRandomVectorScorer when vectorValues is not isntance of HasIndexSlice.

It makes sense to me to add it even just for tests.

[elasticsearchmachine]

💚 Backport successful

Status	Branch	Result
✅	9.2