fix: resolve Python FFI buffer limits and single-quote escaping in to…

[ApurveKaranwal]

Description

This Pull Request addresses three issues identified within the Python FFI bindings and tool-calling parameter parsing logic:

1. JSON Validation for Single-Quoted Parameter Parsing (chat_tools.h): LiquidAI models (LFM2/LFM2.5) format tool calls as Python-like arguments (e.g., my_tool(param='val')). If a parameter value contains an escaped single quote (\'), the parser function py_literal_to_json translated it literally as \' inside the resulting double-quoted JSON string (producing "hello \'world\""). Per RFC 8259, this is an invalid escape sequence. As a result, standard JSON parsers (e.g. C++ picojson and python's json module) threw syntax errors, aborting tool execution. This has been resolved by unescaping \' to ' in the double-quoted JSON output.

2. FFI Document Retrieval Buffer Limit (cactus.py): The vector database supports indexing documents up to 65,535 bytes (Index::validate_documents). However, the Python FFI binding cactus_index_get used a hardcoded buffer limit _INDEX_DOC_BUF_SIZE = 4096. Attempting to retrieve documents exceeding 4KB caused the FFI to return -1, raising a RuntimeError on the Python side. The buffer limit has been increased to 65536 to safely accommodate the database's actual limit.

3. RAG Query Buffer Limit (cactus.py): Increased the query results buffer size in cactus_rag_query from 64KB to 1MB to safely support RAG queries that return multiple large snippets.

---

Verification

• Created a C++ unit test suite at cactus-engine/tests/test_chat_tools.cpp to verify that Python-like literals and LFM2 tool calls containing apostrophes are extracted correctly and parsed successfully by picojson.
• Verified that the test suite compiles and passes successfully.
  Please let me know if you would like any modifications or additions.

Fixes #760