feat: Support S3 Vectors (#3330)

• feat: Support S3 Vectors

Bring Amazon S3 Vectors — AWS’s native, cost-optimised vector store for similarity search, RAG, and AI agents — to AWS SDK for pandas. With this change, indexing a DataFrame and running an approximate- nearest-neighbour query is a one-liner; on-the-fly embedding via Amazon Bedrock is built in.

Implementation lives in a private subpackage awswrangler/s3/_vectors/ and is re-exported flat on wr.s3.*, matching the s3_tables convention.

Public surface (14 functions on wr.s3):

• Buckets: create_vector_bucket, delete_vector_bucket,

        list_vector_buckets, get_vector_bucket

• Indexes: create_vector_index, delete_vector_index,

        list_vector_indexes, get_vector_index

• Data: put_vectors, put_vectors_from_df, get_vectors,

        delete_vectors, list_vectors, query_vectors

Highlights:

• End-to-end RAG: pass text_column + bedrock_model_id to put_vectors_from_df and awswrangler embeds each row via Bedrock (Titan / Cohere) and writes the resulting vectors plus all other columns as filterable metadata. query_vectors mirrors this with query_text / query_vector.
• MongoDB-style metadata filters ($eq,eq,$ne, $gt,gt,$gte, $lt,lt,$lte, $in,in,$nin, $exists,exists,$and, $or) evaluated server-side during search.
• Automatic chunking to AWS API limits (500/put, 100/get, 500/delete) and parallel-segment list_vectors (up to 16 segments).
• Float32 coercion + non-finite rejection; NaN / pd.NA / None metadata cells dropped per row.

• test: add mocked unit tests and live integration tests for S3 Vectors

• tests/unit/test_s3_vectors_mocked.py — 46 tests using unittest.mock, no AWS required. Covers chunking, target resolution, float32 coercion, NaN/pd.NA metadata drop, parallel-segment list, query top-k bounds, Bedrock Titan/Cohere request and response shapes, and aliasing identity.
• tests/unit/test_s3_vectors.py — 10 live integration tests using new vector_bucket (session-scope) and vector_index (function-scope) fixtures in tests/conftest.py. Fixtures self-bootstrap via create_vector_bucket / create_vector_index; no CDK stack required.

• docs: api reference and tutorial for S3 Vectors

• docs/source/api.rst: new “Amazon S3 Vectors” section after “Amazon S3 Tables”, listing all 14 public functions.
• tutorials/043 - Amazon S3 Vectors.ipynb: end-to-end walkthrough covering bucket/index lifecycle, discovery, Bedrock-embedded writes, semantic queries with metadata filters, per-key CRUD, bulk export, and cleanup.
• README.md: tutorial 043 entry.

• fix: type-parametrize np.ndarray for mypy on Python 3.10

• fix: use hyphen in vector index fixture name (underscore is invalid)

• fix: avoid zero-norm vectors in live integration tests (cosine rejects them)

• feat: add chunked parameter to list_vectors

Memory-friendly streaming for indexes too large to materialise in one DataFrame. Mirrors the chunked: bool | int convention used by s3.read_parquet and athena.read_sql_query:

• chunked=False (default) — unchanged; returns a DataFrame and keeps the parallel-segment fan-out (up to 16 segments).
• chunked=True — yields one DataFrame per underlying API page.
• chunked=INTEGER — yields DataFrames of exactly N rows (final frame may be shorter).

Chunked streaming is single-segment and sequential by design; use_threads is ignored in that mode, since lazy iteration across parallel segments would require buffering and defeat the memory win.

Internally, _list_segment is now a thin materialiser around a new _iter_list_pages generator, so the per-segment pagination logic (including max_items enforcement across pages) is shared by both the eager and streaming paths.

Tests: 4 new cases cover per-page yield, exact-size chunking, laziness (no API call before the first next()), and max_items truncation of the chunked stream.