fix(user_memory): prevent LLM no-op explanation from overwriting user profile

[lightzt99]

Summary

• Replace the plain-text extraction contract with a structured JSON response {"changed": bool, "profile": str}, giving the LLM an unambiguous way to express "nothing to extract" without generating natural-language explanations that overwrite the stored profile.
• Wire _call_llm_for_extraction to llm_json_text_with_fallback so models that support response_format={"type":"json_object"} are constrained at the API level; unsupported models degrade automatically.
• Add a three-layer parse fallback so models that ignore structured-output instructions still work: JSON parse → exact-match no-op list → plain-text passthrough.

Motivation

When a conversation contains no extractable profile information, the LLM sometimes returns a verbose refusal (e.g. "整段对话未包含任何可提取的用户信息，保持不变") instead of an empty string. This bypassed the existing exact-match filter and silently overwrote the user's accumulated profile with the refusal text.

A regex guard was considered but rejected: patterns broad enough to catch all no-op phrasing (e.g. 没有.*信息) also match valid profile content such as "用户目前没有任何工作信息，正在求职", producing false positives with no reliable fix. Structured output resolves the ambiguity at the source.

Changes

Area	What changed
src/powermem/prompts/user_profile_prompts.py	Rewrite instructions 6–9 and [Target] section to require {"changed": bool, "profile": str}; clarify JSON keys are always English
src/powermem/user_memory/user_memory.py	_call_llm_for_extraction: use llm_json_text_with_fallback; _extract_profile: replace single-path logic with three-layer parse fallback; drop _NOOP_PROFILE_RE
tests/unit/test_extract_profile.py	14 new unit tests covering all three layers, edge cases, and the core regression scenario

Note: _extract_topics uses a separate JSON contract and is not affected by this change. The inconsistency is known and tracked separately.

Test plan

• pytest tests/unit/test_extract_profile.py — 14 passed
• pytest tests/unit/ — 186 existing tests passed, no regression
• Send a conversation with no user information (e.g. "今天天气好"), confirm existing profile is unchanged in storage
• Send a conversation with valid user information, confirm profile is correctly updated
• Configure a model that does not support response_format (e.g. local Ollama), confirm Layer 3 plain-text fallback activates and logs INFO

Security / privacy

• All three fallback layers log fixed messages only — raw LLM responses are never written to logs, preventing profile content from leaking into log files.
• profile field values are written to storage only; they are not executed as code or commands. The indirect prompt-injection vector (profile content later included in LLM context) is pre-existing and out of scope for this PR.

Fixes oceanbase#933