Enable zero-copy SIMD vector scoring on searchable snapshots (frozen tier)

[ChrisHegarty]

Summary

SIMD-accelerated vector scorers (OSQ / BBQ / int7) previously required an mmap-backed
MemorySegment, meaning searchable snapshot (frozen tier) data had to be copied onto the
heap before scoring. This made quantized vector scoring on frozen indices significantly slower than
on locally mmap'd data.

This PR enables the blob-cache to expose its memory-mapped regions directly to the vector
scorers, eliminating the heap copy and bringing frozen-tier BBQ scoring throughput on par with that
of mmap.

More broadly, this lays the foundation for zero-copy access to blob-cache data beyond
vector scoring. The new DirectAccessInput interface is a general-purpose contract — any
IndexInput consumer can use it to obtain a direct ByteBuffer view of the underlying
data, scoped to a callback with lifecycle managed internally. On the blob-cache side, the
new SharedBytes.IO.byteBufferSlice and CacheFileRegion.tryGetByteBufferSlice APIs are
not vector-specific; they can be used by any future codec or query path that would benefit
from avoiding heap copies when reading from frozen-tier data (e.g. postings, doc values,
stored fields). The FrozenIndexInput, BlobCacheBufferedIndexInput, and
StoreMetricsIndexInput wrappers all propagate the interface, so it is preserved through
FilterIndexInput chains regardless of the consumer.

What changed

New DirectAccessInput interface (libs/core) — a callback-style contract that any
IndexInput can implement to offer a direct ByteBuffer view of its data. The buffer is
scoped to the callback, so all ref-counting and lifecycle is handled internally.

Blob-cache plumbing — SharedBytes.IO.byteBufferSlice and
CacheFileRegion.tryGetByteBufferSlice expose read-only ByteBuffer slices of
memory-mapped cache regions, with ref-count held for the duration of the callback.
FrozenIndexInput implements DirectAccessInput on top of this.
BlobCacheBufferedIndexInput and StoreMetricsIndexInput propagate the interface through
their wrappers so it is not lost by FilterIndexInput chains.

IndexInputSegments.withSlice (libs/simdvec, main21) — a single entry point that
obtains a MemorySegment from an IndexInput by trying, in order:
MemorySegmentAccessInput (mmap) → DirectAccessInput (blob-cache) → heap copy (fallback).
Resource lifecycle is fully internal; callers just receive a MemorySegment in a callback.

Scorer refactoring — the MemorySegment field is removed from all scorer classes
(ES92 int7, OSQ 1-bit / 2-bit / 4-bit). Each scoring method now goes through
IndexInputSegments.withSlice, and the core arithmetic is extracted into static *Impl
methods for clarity. Constructors validate that the IndexInput is a supported type
(MemorySegmentAccessInput or DirectAccessInput), failing fast otherwise.

[elasticsearchmachine]

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

[elasticsearchmachine]

Hi @ChrisHegarty, I've created a changelog YAML for you.