fix: log LLM error response bodies

[RerankerGuo]

Fixes #1775.

Summary

• Include a truncated response body in LLM JSON non-2xx warning logs.
• Add a matching warning log for non-2xx streaming LLM responses before throwing.
• Keep the thrown MemosError details unchanged while making provider rejection details visible in logs.
• Add regression tests for JSON and streaming non-ok response logging.

Validation

• npm test -- --run tests/unit/llm/fetcher.test.ts
• npm run lint
• Attempted: make format; blocked because poetry is not installed in this environment.

[Memtensor-AI]

Automated Test Results: ENVIRONMENT FAILURE

Cloud test-engine rerun against dev-v2.0.22 did not produce a reliable code verdict because the cloud environment failed before executable test phases completed.

• Run: tr-f415ef73-834 on cloud test-engine 10010
• Raw verdict: fail_code_bug
• Summary: Tests failed. Failed cases:

This should be retried after the cloud dependency cache is healthy; do not treat it as proof of a code bug.

[Memtensor-AI]

Automated Test Results: ENVIRONMENT RETRY

The previous cloud run failed before executable test phases because dependency environment preparation did not complete. The shared memos-local-plugin dependency cache is now present on the cloud test-engine host, so this PR has been requeued.

• Retry run: tr-73877358-8e7 on cloud test-engine 10010
• Branch tested: refs/pull/2012/head

I will post the final pass/fail result when this retry completes.

[Memtensor-AI]

Automated Test Results: PASSED

Cloud test-engine rerun against dev-v2.0.22 completed successfully.

• Run: tr-73877358-8e7 on cloud test-engine 10010
• memos_local_plugin/unit: 12 passed, 0 failed, 0 skipped

Manual code review is still required before merge.