perf(agentgateway): parallelize LoB MCP tool listing across fragments

[wangzhengzhuo05]

Fixes #301

What / Why

get_mcp_tools_lob fetched MCP tools from destination fragments sequentially — N fragments × ~3s network round-trip each. Concurrent fetching makes total time ~max(fragment time) instead of the sum (e.g. 14 fragments ≈ 3s instead of ≈44s).

How

• Replaced the sequential for fragment ... await list_server_tools loop with asyncio.gather(*coros, return_exceptions=True).
• URL-missing fragments are still skipped (with warning) at scheduling time; failing fragments are isolated via return_exceptions=True + _log_mcp_server_error — same continuation semantics as before.
• Results are zipped back in fragment order (gather returns in input order), so tool ordering is unchanged.
• Pre-fetch ORD filter, post-fetch names filter, and logging are untouched.

Tests

• test_lists_fragments_concurrently: 3 fragments × 0.05s sleeps complete in < 2× per-call (sequential would be 3×). Mutation check: reverting to a sequential loop makes this test fail (verified).
• test_failure_in_one_fragment_does_not_fail_others: 2nd fragment raises; other fragments' tools still returned.
• test_preserves_fragment_order: completion order is inverted via varied latencies; returned tools still follow fragment order.

Local: 168 passed (test_lob + test_agw_client + test_customer).

AI-generated code disclosure

This change was implemented with the assistance of AI (OpenCode + muse-spark) and reviewed by a human maintainer process; per SAP GenAI guidelines the code has been verified for correctness, security, and license compatibility.