Add a retry mechanism for some requests

[pozdnychev]

The server might answer a code 503 if it's busy. We have to sleep() a bit then retry

[vrancurel]

have to think if retry might be at application level.

[ghost]

Amazon S3 in particular needs a retry if the server returns 503 Slow Down on a PUT.

Droplet blacklists a server if it returns a 5xx HTTP status, which is generally a good thing. But if the application had to retry these errors, it would have to interact with the blacklist logic in strange ways in order to be effective.

I think implementing the retries in Droplet would be cleaner, and all applications would automatically benefit.