feat: improve macOS support and local LLM (GPT4All) compatibility

[lmntrixace]

Issue

Two issues were affecting users on macOS running GPT4All:

1. entities list returned no results even after downloading models. The CLI was searching for models in ~/.cache/ctxgraph/models (a Linux path), while the macOS downloader saves them to ~/Library/Caches/ctxgraph/models. This prevented the extraction pipeline from being loaded.

2. LLM escalation failed: 404 Not Found when pointing CTXGRAPH_LLM_URL at a local server. The code used the provided URL literally (e.g. .../v1) without appending the required /chat/completions path. Additionally, when a local model like GPT4All correctly rejected the "Strict JSON Schema" request format it doesn't support, the fallback to simple prompt-based extraction was never triggered.

Environment

• OS: macOS
• Rust Version: rustc 1.96.0 (stable)
• Local LLM: GPT4All (Reasoner v1 on http://localhost:4891)

Fix

1. Updated find_models_dir in mod.rs to use ModelManager::default_cache_dir() — the same cross-platform path logic the downloader already uses — so the CLI and downloader always agree on where models live.

2. Added ensure_chat_path in llm_extract.rs — a small private helper that appends /chat/completions to any URL missing it, applied across all three env-var loading paths.

3. Changed the extract method from if let Ok to match so any error from the strict JSON attempt (including a 400/404 from an unsupported local model) correctly triggers the simpler fallback extraction method.