Fix: hardcoded embedding dimension

[LaansDole]

Context

Memory embeddings were being stored with inconsistent dimensions.

What's changing

1. embedding.py - Dynamic fallback dimension
   Instead of hardcoding 1536, a _fallback_dim attribute is set to 1536 initially, then updated to the real dimension after the first successful API call:

self._fallback_dim = 1536  # Default; updated after first successful call
# After a successful embed:
embedding = response.data[0].embedding
self._fallback_dim = len(embedding)  # e.g., now 768
return embedding
# All fallback paths now use:
return [0.0] * self._fallback_dim  # matches model's actual dim

2. file_memory.py - Dimension validation at retrieval
   Even if old corrupt data exists, retrieve() now skips mismatched embeddings instead of crashing:

expected_dim = query_embedding.shape[1]
for item in self.contents:
    if item.embedding is not None:
        if len(item.embedding) != expected_dim:  # ← NEW guard
            logger.warning("Skipping memory item %s: dim %d != expected %d", ...)
            continue
        memory_embeddings.append(item.embedding)

3. Makefile - make check-backend for Quality Checks in the future project's consistency