Fix the way environment is handled for worker_tool processes.

Previously the worker process was started with environment variables from whatever first rule happened to start it.

This changes the behavior to:

• allow specifying environment variables via env in worker_tool
• remove step-specific environment variables from the worker launch environment.
• moves $CWD to project root
• moves $TMP to buck-out/bin/<worker_tool target>_worker

This way the worker is started the same way every time. It also avoids including the (potentially huge) environment from the genrule in an actual call to exec.

In terms of code organization, I followed the existing pattern, but I think perhaps it might make sense to move WorkerProcess startup into WorkerTool itself so we don't have to pass so many arguments around.