[dependency] Bump flink to 1.20.5 and 2.1.3

[wenjin272]

Linked issue: #844 #692

Purpose of change

---

Commit 1 — [fix] Keep numpy off the async pool for cross-language RAG queries

Problem

A cross-language RAG query against a Python vector store (VectorStoreCrossLanguageTest) could hang intermittently in CI while passing locally. A Python store's query runs numpy (e.g. ChromaDB normalizes the query embedding via np.array). numpy releases and re-acquires the GIL during that copy, and pemja keeps a single PyThreadState; on the async executor's worker thread the re-acquire can stall, surfacing as a hang on few-core CI runners (machines with spare cores avoid the window). See #844.

Fix

Split the async RAG path so only the numpy step stays on the mailbox thread, leaving the rest async:

• ContextRetrievalAction — for an async Python store, run embed async → normalize synchronously on the mailbox thread → query async with the pre-normalized vector, instead of running the whole query on the worker.
• PythonVectorStore — add embedQuery / normalizeEmbedding / queryNormalized hooks; resolve the embedding model in open(); forward pre-computed vectors to the Python _embedding methods.
• BaseVectorStore — expose getEmbeddingModel() to subclasses.
• ChromaVectorStore + base vector_store.py — add _normalize_embeddings; _query_embedding accepts a pre-normalized ndarray (skips internal numpy) or a raw list.
• Test — update PythonCollectionManageableVectorStoreTest for the Java-side embedding path (pre-computed embeddings + _add_embedding/_update_embedding assertions).

---

Commit 2 — [dependency] Bump flink to 1.20.5 and 2.1.3

Bumps pinned Flink versions: flink.1.20.version 1.20.4 → 1.20.5, flink.2.1.version 2.1.2 → 2.1.3, and aligns flink.2.0.version (2.0.2) and the e2e-integration matrix
accordingly across pom.xml, dist/pom.xml, and the integration e2e pom.

---

Tests

it & e2e

API

no

Documentation

• doc-needed
• doc-not-needed
• doc-included