Pull-thru proxy should not store entire blob in memory

When caching a blob from the upstream registry, the current code will store the entire contents of the blob in memory.

Given that blob sizes can get pretty large, the proxy code should not store the entire blob in memory.

https://hub.docker.com/r/sagemathinc/cocalc/tags can be used to test blob caching with large blobs.

Some (or all?) of the code in the blob uploader might be reused to achieve this. See: https://github.com/quay/quay/blob/f0fb43f44a205f369a9d3ed1558ad456771211e6/data/registry_model/blobuploader.py#L172