fix(pipecat): retain strong references to background storage tasks to prevent GC

[devteamaegis]

Problem

In SupermemoryPipecatService.process_frame, conversation messages are stored
via a fire-and-forget coroutine:

asyncio.create_task(self._store_messages(unsent_messages))

The Task object returned by asyncio.create_task is immediately discarded.
Python's asyncio event loop only holds weak references to scheduled
tasks, so the garbage collector is free to destroy a task at any time — even
while it is still awaiting an I/O call inside _store_messages. On a GC
cycle this can silently kill an in-flight supermemory.memories.add() request,
dropping messages with no error logged.

The SupermemoryCartesiaAgent in the same repo (cartesia-sdk-python) already
uses the correct pattern: it saves each task in a _background_tasks set and
removes it via add_done_callback(discard) once the coroutine finishes.

Fix

Apply the same pattern to SupermemoryPipecatService:

1. Add self._background_tasks: set = set() in __init__.
2. Save every storage task in the set; remove via add_done_callback.
3. Clear the set in reset_memory_tracking().

# before
asyncio.create_task(self._store_messages(unsent_messages))

# after
task = asyncio.create_task(self._store_messages(unsent_messages))
self._background_tasks.add(task)
task.add_done_callback(self._background_tasks.discard)

Tests

tests/test_background_task_tracking.py (all 5 pass):

• test_background_tasks_set_exists – attribute present after init
• test_task_held_during_execution – task is in the set while running, removed after
• test_task_removed_after_completion – set is empty once coroutine finishes
• test_gc_cannot_collect_tracked_task – forced gc.collect() mid-execution can't kill a tracked task
• test_reset_clears_background_tasks – reset_memory_tracking() empties the set