[BUG] RAG text splitter repeats zero-overlap content and exceeds chunk_size

Description

RecursiveCharacterTextSplitter repeats source text with chunk_overlap=0 and can exceed its configured maximum chunk size. Retained separators are also inserted twice during merging.

Reproduction

from crewai_tools.rag.chunkers.base_chunker import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(chunk_size=3, chunk_overlap=0, separators=[""])
print(splitter.split_text("abcdef"))
# Actual: ["abc", "cde", "ef"]
# Expected: ["abc", "def"]

splitter = RecursiveCharacterTextSplitter(chunk_size=5, chunk_overlap=0, separators=["\n", ""])
print(splitter.split_text("aa\nbb\ncc"))
# Actual: ["aa\n\nbb", "\nbb\n\ncc"] (lengths 6 and 7)
# Expected: ["aa\nbb", "\ncc"] (lengths 5 and 3)

Expected behavior

Merging respects both chunk_size and chunk_overlap, without adding separators that are already present. The excess content otherwise changes retrieval chunks and embedding input sizes.

Cause and verification

The overlap loop always retains one split, even for zero overlap, and does not make room for the next split. Separator accounting assumes removed separators even when they were retained. Four regressions fail on current main; a local fix passes all 137 RAG tests, including overlap and separator controls. Python 3.12 on macOS arm64; no LLM calls.

AI assistance was used to investigate, implement and test this change.

Start with RecursiveCharacterTextSplitter in crewai_tools/rag/chunkers/base_chunker.py and inspect its split_text merging behavior for zero overlap and retained separators. Reproduce both examples, then run the RAG test suite; done means all 137 RAG tests pass while chunks respect chunk_size and chunk_overlap without duplicated separators.