Mandatory improvements for production use

[zenria]

Hi,

Here is some improvements to make this crate a bit more production ready:

• do not quit the program if there are too much errors (a lib must never take the responsibility of this)
• honor the buffer_size config option (the config option was there, but never used)
• remove useless Arc<Mutex> in Buffer struct
• keep tcp connection open, reconnect on network errors
• remove various hardcoded thread::sleep()
• remove hardcoded insane 10_000_000 sync_channel bound which eats 2GB (!!!) of RAM because sync_channel always allocate the totality of the buffer, the sync_channel size can now be configured (1000 by default)
• export init_processor so rust-log4rs-gelf create do not have to know what struct to create in order to initialize properly the BatchProcessor
• implement log::Log::flush (flushes the BatchProcessor)
• non blocking logging, even if the buffer is full: a failure in the logging subsystem must not block the application from working
• remove Metronome. metronome logic is implemented with recv_timeout directly in Buffer loop. It seems to be a better option to me than the outer thread filling the buffer with Send events.
• configurable net timeouts

There are still some room for other improvements:

• exponential backoff on reconnection

[cdumay]

Hi,

Great work! Thank you for these improvements which will greatly help to stabilize this library.