fix: ignore tool-call items when collecting user senders

[rendigua2025-gif]

Summary

Fixes #276.

User-side extraction strategies assumed every item in memcell.items exposes a role attribute. Agent/tool-call trajectories can include ToolCallRequest items, which do not expose role, so extract_atomic_facts and extract_foresight could crash before processing otherwise valid user messages.

This PR adds a small shared helper for collecting user sender ids defensively, then uses it in both strategies.

Scope

• Collect sender ids only from items where role == "user".
• Skip non-chat/tool-call items instead of treating them as malformed messages.
• Preserve the existing extraction behavior for normal chat messages.

This does not add tool-call semantics to user-side extraction; it only prevents unrelated tool-call items from crashing the user memory pipeline.

Tests

• tests/unit/test_memory/test_strategies/test_strategy_to_handler_contract.py
• tests/unit/test_memory/test_strategies/test_extract_atomic_facts.py
• tests/unit/test_memory/test_strategies/test_extract_foresight.py

Note: I verified locally on Windows with a test-process-only fcntl stub because the current repository still imports POSIX fcntl during test collection on Windows.