Port pydantic-ai v1.105.0 parity updates for SDK model settings and feature matrix

[Copilot]

This PR advances the v1.87.0 → v1.105.0 port by implementing the relevant SDK-facing change (xAI reasoning effort support) and updating parity tracking for newly introduced upstream items. Python-specific or non-applicable upstream fixes are explicitly marked in the parity table.

• Version tracking

  • Bumped .github/pydantic-ai-version.txt to 1.105.0.

• SDK model settings: xAI reasoning_effort

  • Extended ModelSettings with:
    • reasoningEffort?: "low" | "medium" | "high"
  • Wired the setting into model invocation options via:
    • providerOptions.xai.reasoningEffort

• Tests

  • Added a focused model-settings test to assert reasoningEffort propagation through generation options.
  • Updated type-level coverage in model_settings_test.ts to include the new field.

• Docs / parity matrix

  • Updated docs/concepts/models.mdx to document reasoningEffort.
  • Updated docs/reference/features.mdx with v1.105.0 entries:
    • On-demand capability loading: partial (current deferred tool loading + per-turn instruction resolution)
    • xAI reasoning effort: ported
    • Temporal gateway model-construction fix: not applicable
    • Google cached-content omission fix: not applicable

const agent = new Agent({
  model,
  modelSettings: {
    reasoningEffort: "high",
  },
});