Port pydantic-ai v1.106.0 parity fixes for event stream observers and non-base64 data URIs

[Copilot]

This updates the SDK parity baseline from pydantic-ai v1.87.0 to v1.106.0 and ports the relevant runtime bugfix behavior from that range. The changes focus on stream-handler resilience and data URI decoding behavior in multimodal image output extraction.

• Version alignment

  • Bumped .github/pydantic-ai-version.txt to 1.106.0.

• Multimodal data URI parsing fix

  • Updated extractBinaryImageFromToolOutput() to support both:
    • data:<mime>;base64,<...> payloads
    • valid non-base64 data:<mime>,<percent-encoded...> payloads
  • Added defensive decode handling so malformed payloads return null instead of throwing.

• Event stream handler resilience parity

  • Added regression coverage to ensure runStreamEvents() downstream iteration still completes when an observer-form eventStreamHandler is provided but does not consume the stream.

• Feature parity documentation updates

  • Updated packages/sdk/docs/reference/features.mdx notes for:
    • BinaryImage output behavior (base64 + non-base64 data URI support)
    • ProcessEventStream behavior when observers do not consume events.

// now handled in extractBinaryImageFromToolOutput:
{
  type: "image",
  image: "data:image/svg+xml,%3Csvg%20...%3E%3C/svg%3E",
  mimeType: "image/svg+xml",
}