Draft: Add single-query many-doc bulk scoring API to VectorUtilSupport (#16556)

[CH-Abhinav]

Description

This is Draft PR addressing #16556 and subsequently progressing #15155

This PR is trying to move bulk scoring API down to VectorUtilSupport layer, allowing the PanamaVectorUtilSupport to see bulk operations.

Current progress (float32 dot product):

1. Added zero-allocation void dotProductBulk(float[] query, float[][] docs, float[] results) to VectorUtilSupport.
2. Added scalar fallback loop to DefaultVectorUtilSupport and pass-through to NativeVectorUtilSupport.
3. Added 4 vector unrolling in PanamaVectorUtilSupport

Next steps:

• Update RandomVectorScorer to consume the new VectorUtil.dotProductBulk() method to prove the end-to-end integration.
• Add similar methods to in bulk format.

[CH-Abhinav]

I went ahead and implemented off-heap path to PanamaVectorUtilSupport for dotProductBulk to make end to end connection.

while doing so I did following:-

1. Left the base VectorUtilSupport strictly for on-heap arrays
2. Added a public static SIMD batching method directly to PanamaVectorUtilSupport to handle off-heap vector unrolling
3. Directed existing MemorySegmentBulkVectorOps to new method

If everyone is okay with this implementation pipeline then i will go ahead and implement the remaining similarity functions like cosine and squareDistance . So reviewers out there if you have some time please review this Draft once.

[kaivalnp]

Thanks @CH-Abhinav, this looks like the right path to me.

I think we should also add equivalent functions to NativeVectorUtilSupport though.

[CH-Abhinav]

thanks for the review @kaivalnp I will continue working on PR

But regarding NativeVectorUtilSupport , currently I am by passing FFI to C library and making it always fallback to Panama delegate. If we want proper transfer we will need to write MethodHandle and FunctionDescriptor for bulk Operations including updating underlying C library itself.

So should I do C-library implementations and the FFI calling in this PR or just do SIMD layer and see about native layer later?
I will come up with similary implementations for consine and squareDistance.