Big blobs p2p model sharing example

[IAvecilla]

This is a standalone test harness that exercises the full production P2P model sharing pipeline end-to-end, without requiring actual model data or GPU resources. It simulates the flow that happens during the Warmup phase when new clients join a training run and need to download the current model state from existing peers.

• Sharer peers are initialized with FakeStore (an iroh blob store that generates blobs on-the-fly without holding data in memory), serving both model config and parameter blobs. This allows simulating arbitrarily large models with zero memory allocation.
• Downloader peers use the real production pipeline: blob_ticket_param_request_task for peer selection and ticket requests, NetworkConnection.start_download() for the DownloadManager/iroh Downloader integration, and NetworkConnection.poll_next() for BandwidthTracker and ConnectionMonitor integration.
• Concurrent worker tasks pull from a shared work queue and request blob tickets from peers via PeerManagerHandle (bandwidth + latency-based peer selection with error tracking).