Fix pumper thread leak

[lihaoyi]

Fixes #4184

The basic problem is that we are spawning an PumpedProcessInput from the Mill server's redirected stdin to the stdin of any subprocess we spawn, but the PumpedProcessInput's InputPumper thread only terminates on the source stream closing, and we never close the redirected stdin of the Mill server.

Reading from the shared stdin from multiple tasks/threads concurrently doesn't really make sense, so we should probably not use PumpedProcessInput at all and just pass in a DummyInputStream with zero data. Interactive use cases like consoles and repl are handled by the os.InheritRaw codepath anyway and so won't be affected. stdout/stderr aren't affected, because having multiple tasks/threads writing lines to the same output stream is probably fine (although we could probably do a better job at ensuring it is done at a line-by-line granularity without splitting of lines between threads/tasks)

Tested manually, ran the commands in the linked issue and verified that without this PR the leaked thread can be seen in jstack, but after this PR the leaked thread is no longer present

This might also fix some of the cases where enter-on-rerun for --watch stops worked, perhaps because a leaked InputPumper is grabbing the \n character before it can be read by the Watching.scala logic