[IMP] Make the network retries configurable, with jitter

[hugosantosred]

GitHub rate limits the unauthenticated git operations per IP address. Once the limit is reached it answers with an HTTP 401, which git reports as "could not read Username for 'https://github.com'". A CI job aggregating ~100 repositories from a shared runner hits it regularly, and the 3 retries within 35s of the previous defaults were not enough to get past it.

• Raise the defaults to 4 retries, 15s delay and a 0.2 jitter, so the waits are 15s, 30s, 60s and 120s instead of 5s, 10s and 20s.
• Add the jitter, to keep concurrent aggregations from retrying in lockstep against a rate limiting remote, only to get limited together again.
• Add --retry-max, --retry-delay, --retry-backoff-factor and --retry-jitter, taking precedence over the retry section of the configuration file. It was only configurable per repository in the yaml file, which is not practical in a CI environment with a hundred of them.
• Merge the retry section with the defaults, so setting one parameter no longer drops the others, and raise a ConfigException on unknown ones.
• Document all of it in the README, including the fact that the merge check done while loading the configuration file is not retried on purpose.