fix: preserve complete raw usage objects for xAI Responses (#19657)

## Background

xAI Responses JSON and SSE parsing dropped provider-specific top-level
and nested fields from final-step usage.raw while normalized token
counts remained correct.

## Root Cause

Strict Zod usage schemas stripped unknown fields before conversion to
LanguageModelV4Usage; the immutable reproduction and existing provider
fixtures confirmed data loss at the response parsing boundary.

## Summary

Added JSON-valued catchalls to the xAI Responses usage and token-detail
schemas, updated affected snapshots, and added a patch changeset for
@ai-sdk/xai.

## Testing

Added generate and stream response-path tests covering context details,
server-side tool usage, unknown top-level and nested fields, normalized
counts, last-event-wins streaming behavior, and known-field validation.

## End-to-end Validation

- `replay_original_reproduction`: immutable generate/stream replay
exited successfully with no bug signal.
- `pnpm -C examples/ai-functions exec tsx -e "<live xAI
generateText/streamText web-search probe>"`: both live paths preserved
`context_details` and `server_side_tool_usage_details` in final-step
`usage.raw`.

## Related Issues

Fixes #19652

Closes #19656

---------

Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com>
Co-authored-by: gr2m <39992+gr2m@users.noreply.github.com>