feat: support configurable safety_settings on ChatVertexAI

[pinetops]

Summary

ChatGoogleAI exposes a safety_settings field passed through to the API as safetySettings, but ChatVertexAI has no equivalent — callers can't set Vertex AI harm-category thresholds, so the provider defaults always apply.

This adds a safety_settings field to ChatVertexAI (default []), mirroring ChatGoogleAI:

• cast it in @create_fields
• conditionally add it to the request body as safetySettings

Behaviour

Default [] is a no-op (an empty list is not added to the request), so existing behaviour is unchanged. Callers who need to tune harm thresholds — e.g. a stricter or more permissive filter for their domain — can now do so:

ChatVertexAI.new!(%{
  model: "...",
  endpoint: "...",
  safety_settings: [
    %{"category" => "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold" => "BLOCK_ONLY_HIGH"}
  ]
})

Test

Mirrors the existing ChatGoogleAI safety-settings tests: present settings are sent as safetySettings; an empty list is omitted. chat_vertex_ai_test.exs passes (41 tests).

[brainlid]

Hi @pinetops! I see the PR is still a draft. Please let me know when it's ready. Thanks!