Abstracting away SubcommandManager's signal handling

[filipcacky]

Currently, SubprocessManager and by extension Runner and Deployer hook an interrupt signal handler via either signal or asyncio. This is slightly problematic when integrating metaflow with an existing codebase, which already does its own signal handling, because both registering a signal with either asyncio or signal overrides any previously set signal handler.

I propose abstracting away the hook in SubprocessManager to a new class SignalManager and allowing users to supply their own instance of it to Deployer and Runner. This would be a non-breaking change.

Adding a POC impl in PR.