feat: add multimodal image support for vision-capable LLMs

[lcolok]

Summary / 概述

Enable channels (e.g. Telegram) to forward user-attached images to vision-capable LLMs. Previously, images were downloaded and stored locally but never passed to the provider — the media pipeline was broken at the agent layer (msg.Media was discarded, BuildMessages received nil for media, and Message had no image field).

让 channel（如 Telegram）能将用户发送的图片转发给支持视觉的 LLM。此前图片虽已下载存储到本地，但从未传递给 provider — 媒体管道在 agent 层断裂（msg.Media 被丢弃，BuildMessages 的 media 参数硬编码为 nil，Message 也没有图片字段）。

Changes / 改动

File / 文件	Change / 改动
`protocoltypes/types.go`	Add `ImageURL` type, `ContentBlock.ImageURL` field, `Message.ContentParts` field / 新增 `ImageURL` 类型及相关字段
`providers/media.go`	New file — `LoadMediaAsContentParts()`: local images → base64 data URL (jpg/png/gif/webp, ≤5MB) / 新建 — 图片加载工具函数
`agent/loop.go`	Add `Media` to `processOptions`, resolve `media://` refs to local paths, pass to `BuildMessages` / 解析媒体引用并传递
`agent/context.go`	`BuildMessages` attaches `ContentParts` (text + image blocks) when media is present / 有媒体时构建多模态内容
`openai_compat/provider.go`	`Content` field becomes `any` (string or `[]contentPart`), following OpenAI Vision API spec / 适配 OpenAI Vision API
`anthropic/provider.go`	Map `ContentParts` to `NewImageBlockBase64()` via `parseDataURL()` / 适配 Anthropic Vision API

Design / 设计

• Minimal & upstream-friendly: Only extends existing types, no breaking changes / 最小改动，仅扩展不破坏
• Follows OpenAI Vision API spec: `content` field is `string | []ContentPart` / 遵循 OpenAI Vision API 规范
• Backward compatible: When no media is attached, all code paths remain unchanged. `Content` (string) is always preserved as fallback / 无图片时所有行为不变
• Both providers supported: OpenAI-compatible (content array) and Anthropic (image block base64) / 两种 provider 均已适配

Data Flow / 数据流

Channel downloads image → MediaStore.Store() → media://ref
  → processMessage: Resolve(ref) → /tmp/xxx.jpg
    → processOptions.Media → BuildMessages
      → LoadMediaAsContentParts → data:image/jpeg;base64,...
        → ContentParts on user Message
          → OpenAI: []openaiContentPart (Vision API format)
          → Anthropic: NewImageBlockBase64()

Test Plan / 测试计划

• Compiles with `CGO_ENABLED=0 go build ./cmd/picoclaw/...`
• Telegram: send image + text "这是什么" → AI describes the image
• Telegram: send image only → AI describes the image
• Telegram: send text only → normal reply, no regression
• Send non-image file → no crash, text part processed normally

[lcolok]

Closing in favor of #1020 which was merged with a more complete implementation (streaming base64, filetype detection, configurable size limit).

One piece that #1020 doesn't cover: Anthropic native provider vision support (anthropic/provider.go). I'll submit that as a separate, focused PR.