Add embedder auto-truncation and settings serialization

[bond005]

• EmbedderOpenAI now auto-truncates input to Settings.embedder_token_limit tokens
  before every API call (per-instance overrides via constructor). The context_truncator
  plumbing was removed from Index / KnowledgeGraph — truncation now lives inside the embedder.

• Search engines no longer accept max_context_length / tokenizer_backend /
  tokenizer_model; BaseEngine reads these from Settings.

• TokenTruncation (local backend) now accounts for special tokens ([CLS]/[SEP]),
  fixing silent over-limit requests against BGE/E5-style models.

• Settings.save(path) / Settings.load(path) — JSON serialization of the config
  (Literal/int/str validation; storage_folder deliberately excluded) for reproducible runs.

• Centralized tokenizer & token-limit configuration in the Settings singleton:
  new typed fields tokenizer_embedder_backend, tokenizer_llm_backend,
  tokenizer_embedder_name, tokenizer_llm_name, embedder_token_limit,
  llm_token_limit, llm_context_token_limit (with Literal["tiktoken","local"] validation).

• Bugfix: SQLite database disk image is malformed corruption resolved via sqlite_synchronous=2.

• New example examples/local_embedder_with_short_context.py (local vLLM embedder, 512-token context);
  README, AGENTS.md and subpackage READMEs updated.