fix(python-sdks): handle pydantic search results in memory dedup helpers

[vivekvar-dl]

Fixes #1266

Problem

The three SDK-based Python packages pass search results into dedup/format helpers written for plain dicts, but typed SDK results are pydantic models — attribute access, snake_case fields, no .get():

• pipecat / cartesia: deduplicate_memories's unique_search calls r.get("memory", "") → AttributeError on a model. The exception is swallowed upstream (generic frame/event handlers), so no memories are injected for any user whose lookup returns search results — it only appears to work for brand-new users.
• agent-framework: extract_memory_text handles only dict and str, so model items fall through to None and every search-result memory is silently dropped.

A version wrinkle makes this worse: on current supermemory (3.56.0) the profile endpoint types search_results.results as List[object], so items arrive as camelCase dicts — while typed models (e.g. search_memories_response.Result) expose snake_case attributes. The packages pin supermemory>=3.16.0, so both shapes occur in the wild and the helpers must accept both.

Fix

Implements the shared field-extraction approach described in the issue:

• pipecat + cartesia utils.py (kept byte-identical, matching the existing duplication between the two packages): a private _get_result_field(result, *keys) reads a field from either shape — dict lookup with camelCase/snake_case keys, or attribute access. Used in unique_search and format_memories_to_text. Plain-string items and camelCase dicts behave exactly as before; updated_at/updatedAt both resolve for the timestamp.
• agent-framework utils.py: extract_memory_text gains an attribute-access fallback for non-dict, non-str items. Dict/str/None behavior is unchanged.

Tests

• New tests/test_utils.py in pipecat and cartesia (unittest style, matching the existing suite): model items survive dedup, dedup against profile strings, camelCase dicts unchanged, timestamp formatting for both shapes, plain strings verbatim. The dependency stubs moved from test_empty_profile.py into a shared tests/conftest.py so both test modules reuse them.
• Three new cases in agent-framework's existing TestDeduplicateMemories (pytest style): model items extracted, deduped against static, empty/None memory filtered.

All three suites pass locally (Python 3.12): pipecat 9/9, cartesia 9/9, agent-framework 57/57 (with agent-framework-core==1.0.0rc3 — the current stable releases break the package's import, tracked separately per the issue's closing note).

Also verified the issue's exact reproduction against the real SDK (supermemory 3.56.0): search_memories_response.Result now dedups and formats correctly in all three packages, and camelCase dict inputs are unchanged.