Fix intermittent empty/truncated Grok search responses

[dztabel-happy]

Summary

This fixes an intermittent issue where web_search can return an empty or truncated content string even though the upstream request succeeds and sources_count is non-zero.

Reproduction

Using the current main implementation against a real Grok-compatible endpoint, repeated calls to GrokSearchProvider.search(...) intermittently returned incomplete results:

• empty string responses
• very short partial prefixes like "**" or "**Scrape.do** is a"
• normal full responses on other runs

In one 6-run sample before this patch, 2 runs returned an empty string and several others were truncated.

Root Cause

The provider currently depends on streaming chat completions (stream=True) and concatenates SSE deltas.
For this upstream, streaming responses are not consistently delivered as a complete answer, which causes partial or empty content to propagate into web_search.

Fix

• switch Grok provider search/fetch/URL description/source ranking calls to non-stream completions (stream=False)
• add explicit Accept and User-Agent headers for the upstream API request
• parse standard JSON chat completion bodies first
• keep an SSE-text fallback parser for compatibility if an upstream still responds with data: lines

Validation

• Added regression tests for:
  • non-stream payload + request headers
  • JSON completion parsing
  • SSE-text fallback parsing
• Ran uv run --extra dev pytest tests/test_grok_provider.py -q
• Re-ran the same 6-run live search loop after the patch: EMPTY_COUNT dropped from 2 to 0
• Verified end-to-end MCP web_search calls via local stdio server returned non-empty content across repeated runs