fix: force batch_size=1 for gemini-embedding-2 models

[Maanik23]

Summary

Fix GeminiEmbedder.create_batch() silently returning fewer vectors than inputs when using gemini-embedding-2-preview or gemini-embedding-2 models.

Closes #1467

Problem

The gemini-embedding-2* models do not support batching via embed_content(contents=[...]). They treat the list as parts of a single document and return only one embedding vector regardless of how many strings are passed.

This caused create_batch() to silently return fewer vectors than inputs, leading to ValueError in downstream zip(..., strict=True) calls during entity deduplication (e.g. in _semantic_candidate_search).

Fix

1. Extend the batch_size=1 special case — the same defense already applied to gemini-embedding-001 is now applied to the -2 and -2-preview models. This ensures each input is sent individually, matching the API's actual behavior.

2. Add a defensive length-mismatch check — if a future model (or misconfiguration) causes the API to return fewer embeddings than requested, create_batch() now detects this and triggers the existing per-item fallback path instead of silently returning partial results.

Tests

Added 6 new tests covering:

• gemini-embedding-2-preview defaults to batch_size=1
• gemini-embedding-2 defaults to batch_size=1
• gemini-embedding-001 still defaults to batch_size=1 (existing behavior)
• Other models use default batch size (100)
• Explicit batch_size overrides model-specific default
• Length-mismatch triggers fallback to individual processing