Account for image inputs in LLM gateway token estimates

Problem

The LLM API gateway does not estimate image input tokens from image dimensions
and the target model's visual processor. Multimodal requests can therefore
reserve input-token capacity based on serialized payload size rather than the
visual tokens consumed by inference.

Desired behavior

• Read dimensions from base64-encoded image headers without decoding pixels.
• Bound header inspection so malformed images cannot cause unbounded scans.
• Apply model-specific patch, tile, and fixed-token estimators where known.
• Use a conservative fallback for unknown models or unavailable dimensions.
• Keep model-name matching explicit so unrelated names cannot select a cheaper
  estimator.

Acceptance criteria

• Chat Completions and Responses admission accounting includes every image.
• Common NIM vision models use documented processor equations.
• Unknown models receive a size-based bounded estimate.
• PNG, JPEG, GIF, and WebP base64 inputs are covered by tests.

Start by locating the admission-accounting paths for Chat Completions and Responses and the existing token-estimation and model-name matching logic. Trace how image inputs are represented, then use the acceptance criteria to define completion: bounded header inspection, documented estimators for common NIM vision models, conservative unknown-model handling, and tests for PNG, JPEG, GIF, and WebP base64 inputs.