Summary

• add a small aggregate-memory example to the profiler guide
• show how to pair Accelerator.profile() with torch.cuda.reset_peak_memory_stats()
• document a lightweight way to log peak allocated and reserved memory without adding a new manager API

Why

Issue #3300 asks for a memory logging manager, but the maintainer discussion suggests that a documented minimal example may already cover the main need for many users.

Testing

• docs-only change
• reviewed rendered markdown locally in diff form

Related

• Closes [Feature Request] add a memory logging manager #3300