feat(xai): add mcpServer tool for MCP server support

[boylec]

Summary

This PR adds MCP server tool support and aligns xAI streaming behavior with the OpenAI provider.

Changes

1. MCP Server Tool Support

• Adds xai.tools.mcpServer() for connecting to remote MCP servers via xAI's Responses API
• Includes MCP type definition in XaiResponsesTool with all xAI API fields
• Args validation and camelCase→snake_case transformation
• Streaming event schemas for MCP calls (mcp_call, response.mcp_call.* events)

2. Streaming Refactor (Matches OpenAI Pattern)

The xAI Responses API sends text in two places:

1. Incremental deltas via response.output_text.delta events
2. Complete text in the response.output_item.done event

Previous behavior: Emitted text-delta from both, causing duplicate text in the stream.

New behavior (matches OpenAI):

• response.output_text.delta → emit text-delta (streaming content)
• response.output_item.done → emit text-end (signal completion, no content)

This aligns with packages/openai/src/responses/openai-responses-language-model.ts which emits text-end from the done event rather than re-emitting text content.

Test plan

• Unit tests for prepareResponsesTools MCP handling (4 new tests)
• Manual testing with real MCP server (Google Places + PostGrid)
• Verified streaming works with streamText() and Output.object()
• All 112 xAI package tests pass

Usage

import { xai } from '@ai-sdk/xai'
import { streamText } from 'ai'

const result = streamText({
  model: xai.responses('grok-4-1-fast-reasoning'),
  tools: {
    web_search: xai.tools.webSearch(),
    my_mcp: xai.tools.mcpServer({
      serverUrl: 'https://my-server.com/mcp',
      serverLabel: 'my-tools',
      allowedTools: ['tool1', 'tool2'],
    }),
  },
})

[boylec]

Please label as ai/provider