Support image input for local on-device models

Problem

Vision support was added for cloud providers (OpenAI-compatible, Anthropic) in !86 (merged). Local on-device models (Gemma 4 on the LiteRT-LM runtime) do not currently support image input: this was explicitly deferred rather than implemented, per review discussion on !86 (merged). Users on the local model can't share images in chat the way cloud users can.

Proposed solution

Investigate whether the current LiteRT-LM runtime and Gemma 4 E2B model build accept multimodal input, then either:

• Give local models the same staged-image workflow as cloud models, converting staged images into whatever session format the runtime expects, or
• Conclude support isn't currently feasible and document why, updating the "Provider support" table in docs/src/conversation/image-input.md.

Alternatives considered

Deferral was the deliberate choice in !86 (merged); this issue is the follow-up. If the runtime can't do vision yet, documenting the gap is the acceptable outcome.

Additional context

• docs/src/conversation/image-input.md: "Provider support" table
• src/eigin_ios/Sources/Inference/Clients/Local/: local client, Gemma4/GemmaRuntime.swift is where session input is assembled
• !86 (merged): cloud vision implementation and the deferral discussion