Add `add_many` for batch embedding

[himanshu231204]

Problem

add() embeds one string at a time. fastembed's embed() takes an iterable and batches internally, so storing 1,000 memories one by one pays per-call overhead a thousand times over.

This blocks bulk import (#20) from being usable at any real size, and makes the benchmark harness (#21) slow to set up.

What to build

MemoryStore.add_many(items: Sequence[MemoryInput]) -> list[Memory]

One embedding call for the whole batch, one transaction for the inserts.

Acceptance criteria

• Single embed() call for the batch, not one per item
• Single transaction — a partial batch on failure is worse than none
• Returns stored Memory records in input order, with their ids
• Same validation as add(): empty content rejected, tags normalized
• Configurable batch size, so a 100k import doesn't try to embed everything at once and exhaust memory
• Tests: ordering preserved, all records findable afterwards, empty input handled, rollback leaves the database unchanged
• Benchmark showing the improvement over looping add(), posted on the issue
• Docs in docs/guide/python-library.md plus CHANGELOG.md

Notes

Deliberately library-only for now — no MCP tool. An agent storing memories one at a time is the normal case, and a bulk tool invites an agent to dump a conversation into memory wholesale, which is exactly the low-value noise that degrades search. Bulk import is a human operation.

Pointers

• src/localmem_mcp/store.py — add(); note self.embedder.embed() already accepts a sequence
• The _lock and with self._conn transaction pattern in add()