fix(grok): preserve Responses vision tool output images

[251392463]

Summary

• preserve images sent by Grok Build in function_call_output.images
• extract input_image parts from structured function_call_output.output arrays
• keep function_call_output.output as text while lifting media into a following user message in call order
• cover parallel tool outputs and invalid/empty image payloads with regression tests

Root cause

Grok Build 1.0.5 can return local read_file results as either top-level images[] or structured output[] parts. The Grok Responses sanitizer previously reduced every tool result to a string and silently discarded those images (or serialized image data as ordinary text). The next continuation therefore reached xAI without the vision context that the client had supplied, which can lead to an unrelated continuation as described in #6185.

This change keeps the xAI-compatible textual tool output and sends the extracted images as standard Responses input_image content immediately after the consecutive tool outputs, preserving per-call attribution and order.

Validation

• go test ./internal/service -run 'TestSanitizeGrokResponsesModelInput|TestPatchGrokResponsesBody.*Images' -count=1
• go test -tags=unit ./internal/service ./internal/pkg/apicompat -count=1
• go vet -tags=unit ./internal/service

Fixes #6185

[github-actions]

All contributors have signed the CLA. ✅
_{Posted by the CLA Assistant Lite bot.}

[251392463]

I have read the CLA Document and I hereby sign the CLA

[251392463]

公司内部环境验证，问题已经解决了