Add Smarter Backoffs for MPI ring connection

[mhamilton723]

Hey, I am running many LightGBM training runs as part of LightGBM on Spark's CI process and i notice that it does alot of waiting because of bad network and this dramatically slows the tests:

LightGBM/src/network/linkers_socket.cpp

Line 190 in 0a4a7a8

const int connect_fail_delay_time = 10 * 1000; // 10s

Could you potentially make it first wait for half a second, then wait for 10 seconds? Or make this configurable. Thanks for the help!

[guolinke]

This could be configurable.
I will create a PR for it.

[mhamilton723]

Thank you @guolinke for considering this!