Add Float16VectorSimilarityQuery for similarity threshold search over fp16 vectors

[rcpeken]

Description

Part of #16725. Adds Float16VectorSimilarityQuery, the fp16 counterpart of
FloatVectorSimilarityQuery / ByteVectorSimilarityQuery, so similarity
threshold search works on fields indexed with KnnFloat16VectorField.

The class mirrors FloatVectorSimilarityQuery on top of the fp16 APIs added in
#16383 (Float16VectorValues, LeafReader#getFloat16VectorValues, the
short[] searchNearestVectors overload, VectorUtil#checkFiniteFloat16).
toString converts the first target component with Float.float16ToFloat so
it prints a readable value rather than the raw bits.

One deliberate difference: only the Adaptive variant is provided. In the
float and byte queries, Explicit is @Deprecated and exists to stay
equivalent to the Lucene 10.4 constructors; fp16 has no such legacy callers, so
adding a new API that is deprecated from day one didn't seem useful. The class
keeps the same sealed shape, so Explicit can be added later if reviewers
prefer parity.

Tests

TestFloat16VectorSimilarityQuery plugs fp16 types into the shared
BaseVectorSimilarityQueryTestCase. Ran it with -Ptests.iters=20, and with
-Dtests.monster=true so testApproximate also exercises the HNSW path; all
pass. tidy and ecjLint are clean.

[rcpeken]

The failing arm CI job is in :lucene:analysis.tests:test, which this PR doesn't touch (the change is limited to lucene/core search). The previous run on the same code passed on all platforms, so this looks like an unrelated flaky test. Could someone re-run the job when convenient?

[kaivalnp]

LGTM, thanks!

Edit: I will merge this in a day, unless there are further comments. The test failure does seem unrelated, I opened #16736. The same test failed on a re-run, either we got really (un)lucky or the failure is very common.