API gateway token counting is accurate and media-aware

What changed

The gateway's token-count endpoints now estimate against the converted request the downstream provider actually receives, instead of a rough text-only walk of the raw body:

• POST /v1/messages/count_tokens (Anthropic dialect): counts with per-dialect text tokenizers (real o200k BPE for OpenAI-style targets), pixel-based image costs, per-dialect audio/video rates, and the tool definitions as they reach the wire. When the target is an Anthropic-dialect endpoint, the provider's own authoritative count_tokens is preferred, with a fast local fallback.
• Gemini POST /v1beta/models/{model}:countTokens: requests carrying media (inlineData/fileData) now return HTTP 200 with a media-inclusive estimate. Previously they were rejected with 400 INVALID_ARGUMENT, forcing the client to fall back to its own local guess.

Why this matters to the user

Agent/CLI sessions through the gateway (e.g. Claude Code, gemini-cli) rely on these endpoints for context tracking. With accurate counts — including images and other media — auto-compaction triggers at the right time instead of overflowing the provider's context limit (issue #17079) or compacting far too early.

What the user should do

Nothing — automatic. No settings change.

Notes for release manager

Forwarding tool-result images as real image content (relocation into user file parts) landed on main separately and is not part of this change; this PR's user-visible change is the counting. Clients that depended on the Gemini 400-on-media behavior now receive a 200 estimate instead.