Limit concurrent downloads in model sharing

[IAvecilla]

Currently, we’re limiting the concurrent blob requests for each model parameter during P2P sharing using the max_concurrent_parameter_requests variable. This only limits the requests for blobs, not the actual downloads. That might be contributing to network saturation during the P2P model sharing process.

It could be worth exploring a limit on total concurrent downloads (not just requests). We might also tweak the current value to find a better trade-off between keeping the process within a reasonable time and avoiding network saturation from too many simultaneous downloads.