[Meta] collection of log improvements

[kennytm]

This issue collects some quality issues we've identified when trying to debug a problem using the log files alone. Feel free to edit and add more items.

• 1. In BR, an error due to context.Canceled should not produce any log above DEBUG level. This is just visual clutter and mask the real error.
• 2. In BR, logs of errors from TiKV should include the IP address of the TiKV, so that we don't need to fetch the log from every node in the cluster.
• 3. In TiKV, the download step should record how long the process has taken (whether success or not).
• 4. In BR, retryable errors should not result in an [ERROR] level log.
• 5. In BR restore, the "Full restore" is printed too frequently (0.2s/log?!), drowning out useful information. utils: fix excessive refresh frequency when BR_LOG_TO_TERM=1 #334
• 6. In BR, ensure all keys are formatted using upper-case hex. We can wrap keys in kv.Key to get the proper format. Printing []byte using zap.Binary will get base64 output (wrong) while using zap.Reflect will get you even worse stuff like "t\\u0000\\u0000…. Print keys on log as hex string #359

[YuJuncen]

I will try to do (6) after #266 merged. (Because #266 make a big change, fix this before it merges might do some extra work and make things more complex...)