Add Alternator vector search support

[dkropachev]

Replaces #87.

This carries the vector search support from #87, scoped to the vector API/client integration work:

• Add com.scylladb.alternator.vectorsearch request/response support for VectorIndexes, VectorIndexUpdates, VectorSearch, FLOAT32VECTOR, and query scores.
• Auto-register VectorSearchInterceptor in the sync and async Alternator client builders.
• Add unit coverage for vector JSON conversion, sync/async response transformation, async query result extraction, checksum/header handling for modified vector responses, and vector integration helpers.
• Add integration tests that skip cleanly when the target Scylla/Alternator instance does not support vector indexes or has vector store disabled.

Split out of this PR:

• Demo2 shutdown cleanup: separate PR.
• Live-nodes polling client shutdown cleanup: separate PR.
• Compression/vector interceptor chaining: separate stacked PR based on vector-1.

Tests:

• JAVA_HOME=/usr/lib/jvm/java-11-openjdk-amd64 PATH=/usr/lib/jvm/java-11-openjdk-amd64/bin:$PATH mvn -q fmt:check
• JAVA_HOME=/usr/lib/jvm/java-11-openjdk-amd64 PATH=/usr/lib/jvm/java-11-openjdk-amd64/bin:$PATH mvn -q -Dtest=VectorSearchInterceptorTest,VectorSearchHttpInterceptorTest test
• JAVA_HOME=/usr/lib/jvm/java-11-openjdk-amd64 PATH=/usr/lib/jvm/java-11-openjdk-amd64/bin:$PATH mvn -q test

[m-szymon]

Comparing to ResponseCompressionInterceptor there is no recording of cancellation. AI says that ResponseCompressionInterceptor.DecompressingSubscriber does it correctly:

private boolean cancelled;

@Override
public void onNext(ByteBuffer byteBuffer) {
    if (isCancelled()) { return; }       // ← stops buffering
    ...
}

@Override
public void onComplete() {
    ...
    if (!isCancelled()) {
        downstream.onComplete();          // ← only emits if not cancelled
    }
}

@Override
public void cancel() {
    cancelled = true;                     // ← records cancellation
    upstream.cancel();
}

[m-szymon]

GzipRequestInterceptor doesn't set Content-Length. AI says that SDK fills in Content-Length, but only when it's absent. So now in case of VectorSearch + compression Content-Length is wrong. Either all interceptors should set it or none.

[m-szymon]

Regarding conversions of FLOAT32VECTOR consider also some kind of compatibility switch in client.

Default: Keep the current behavior.
Server FLOAT32VECTOR -> interceptor -> DynamoDB L of N values
Applications can use: List<AttributeValue> values = item.get("embedding").l();

Opt-in: preserve mode
Expose server vectors as the existing magic-prefixed B attribute:
Server FLOAT32VECTOR -> interceptor -> magic-prefixed DynamoDB B
This is the same local representation created by: Float32Vector.toAttributeValue(0.25f, -1.5f, 3.0f)
It permits lossless read-modify-write:

Map<String, AttributeValue> item = client.getItem(request).item();
client.putItem(PutItemRequest.builder().tableName("copy").item(item).build());

In preserve mode, embedding.l() will not not work for optimized vectors.
Instead, Float32Vector should offer representation-independent accessors:

float[] values = Float32Vector.toFloats(item.get("embedding")); // decodes B-magic or converts List of Numbers
List<AttributeValue> numbers = Float32Vector.toNumberList(item.get("embedding")); // directly returns List of Numbers or converts B-magic

Both forms can be distinguished explicitly: boolean optimized = Float32Vector.isFloat32Vector(embedding);