feat(voyage): add voyage-context-4 and voyage-4 series models

[fzowl]

Summary

Adds voyage-context-4 and the rest of the current Voyage AI model lineup to the cost/context maps, and fixes contextual embedding input handling.

Models added (both model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json)

Model	input_cost/token ($/1M)	ctx	mode
voyage/voyage-context-4	1.2e-07 ($0.12)	120000	embedding
voyage/voyage-4	6e-08 ($0.06)	32000	embedding
voyage/voyage-4-large	1.2e-07 ($0.12)	32000	embedding
voyage/voyage-4-lite	2e-08 ($0.02)	32000	embedding
voyage/voyage-4-nano	0.0	32000	embedding
voyage/voyage-multimodal-3.5	1.2e-07 ($0.12)	32000	embedding

voyage-4-nano is an open-source model that runs locally (loaded by the voyageai Python package). Cost 0.0, consistent with other locally-run models in the map (e.g. ollama/*). Metadata (context length, dimensions) still useful.

Contextual embedding input

VoyageContextualEmbeddingConfig.transform_embedding_request now passes list[str] straight to the contextualized embeddings API, and only forwards list[list[str]] when the caller actually provides that nested shape (bare str normalized to [str]).

Tests

• voyage-context-4 added to contextual model detection test
• New test_contextual_embedding_request_input_normalization (str / list[str] / list[list[str]])
• tests/llm_translation/test_voyage_ai.py: 16 passed, 1 skipped

Double-check

• Whether voyage-multimodal-3.5 should carry a per-pixel cost (mirrors existing multimodal-3, token cost only)