perf: deduplicate identical texts within a single embed_batch() call

[ztsalexey]

Context

Deferred from PR #235 — Copilot flagged redundant API calls for duplicate texts, and we replied it's a valid but low-priority future optimization.

Problem

If embed_batch(&["foo", "bar", "foo"]) is called with duplicates and none are cached, miss_texts includes "foo" twice, resulting in a redundant API call for the same content. The cache still returns correct results — it just wastes an HTTP round-trip for the duplicate.

Proposed Solution

Before calling the inner provider, group misses by cache key:

1. Build a HashMap<String, Vec<usize>> mapping unique text → list of original indices
2. Call the inner provider only for unique texts
3. Fan out the returned embeddings back to all original indices

Why It Was Deferred

• Cache works correctly as-is — the redundant call is a minor inefficiency
• In practice, callers rarely pass duplicate texts in a single batch
• The dedup + fan-out logic adds complexity for minimal real-world savings

Effort

Medium — need the dedup map, unique-only provider call, and index fan-out.