fix(storage): stop ChromaDB from crashing when reopening an existing …

[Legion345]

What does this PR do?

Fixes #1089.

When opening an existing palace, get_or_create_collection was called with hnsw metadata on every open. In ChromaDB 1.5.x, if that metadata differs from what's already stored, the Rust bindings segfault — no traceback, just a core dump. This is what causes the stop hook to crash at session end.

Fix: try get_collection first, fall back to create_collection only when the collection doesn't exist yet. Existing palaces open without touching their metadata; new ones are created with the full settings as before.

Credit to @jphein who described this exact approach in #1089 and has been running it on their fork since April 10 with zero crashes across 400+ starts.

How to test

Open a palace twice in the same session:

from mempalace.backends.chroma import ChromaBackend
import tempfile, os

with tempfile.TemporaryDirectory() as d:
    palace = os.path.join(d, "palace")
    backend = ChromaBackend()
    backend.get_collection(palace, collection_name="mempalace_drawers", create=True)
    backend.get_collection(palace, collection_name="mempalace_drawers", create=True)
    print("no crash") 

Or run the new regression tests directly:
python -m pytest tests/test_backends.py -v -k "idempotent or preserves_existing"

Checklist

• Tests pass (python -m pytest tests/ -v)
• No hardcoded paths
• Linter passes (ruff check .)