Allow shell evaluation on model definition

[kyujin-cho]

Main idea

As for now, Backend.AI only supports passing command line arguments on model definition's "command" directive. We can think of expanding the feature by switching under-the-hood implementation at the kernel runner to treat whole arguments passed as a shell command.

Alternative ideas

No response

Anything else?

No response

[achimnol]

If you want to join the string pieces when the startup command is given as a list to pass the join result to a shell, use shlex.join().