feat(tokens): account for image inputs in token usage

[chaliy]

What

Image inputs now contribute an approximate token cost to reported usage, across every endpoint that accepts images:

• Chat Completions → prompt_tokens
• OpenAI Responses API → input_tokens
• OpenResponses → input_tokens

Follow-up to #73 (which made the endpoints accept and gate multimodal input). This makes usage reflect images instead of silently ignoring them.

Why

After #73, image parts were accepted and validated but dropped before token counting, so a request with images reported the same token usage as a text-only request. Clients that budget against usage (cost estimation, rate-limit simulation) saw no signal from image content.

How

• New estimate_image_tokens(detail) in src/tokens.rs (+ IMAGE_TOKENS_LOW/IMAGE_TOKENS_HIGH constants), exported from the crate root.
• Decision: the simulator never fetches or decodes image bytes, so the real tile-based count (which needs pixel dimensions) is impossible. We approximate from the detail hint using OpenAI's documented gpt-4o costs: "low" → 85 tokens; "high"/"auto"/unset → 765 (a representative ~1024×1024 image, 85 + 4*170).
• Chat Completions count_request_tokens adds per-image cost from each part's detail.
• count_responses_input_image_tokens (Responses) and count_openresponses_input_image_tokens (OpenResponses) fold image cost into input-token counts. The OpenAI Responses input_image part has no detail, so it uses the high-detail default.
• Image content still does not influence the generated output text.
• Specs (api-endpoints.md, responses-api.md) and docs (docs/api.md) updated.

Risk

• Low
• Only affects reported token counts when image parts are present; text-only requests are unchanged. Approximation is documented as such. Verified end-to-end: high-detail image adds 765, low-detail adds 85 across all three endpoints.

Checklist

• Tests added or updated (estimate_image_tokens; per-endpoint image-token accounting)
• Backward compatibility considered (text-only usage unchanged)

---

Generated by Claude Code