fix(daemon): plumb through flush watch errors (#5070)
### Description
First pass as plumbing the status of watching the flush directory to the
glob watcher and watcher configs.
Notable changes:
- Before we start to process flush commands we wait until we get
confirmation that we are watching the flush directory so we can properly
handle flush commands
- Failing to watch the flush directory will cause the watch future held
by the server to exit. The daemon server won't shut down, but will
respond with errors to all glob watching related RPC requests
Reviewer notes:
- The current timeout logic means that if erroring on setting up the
flush directory takes over 500ms we skip the flush and proceed as usual
- We now return errors for `get_changed_outputs` RPC call
- I'm unsure of the changes to the `serve` function. I've manually
tested them, but they feel off. In the future when we possibly have
multiple long running futures this won't scale.
### Testing Instructions
Unit test should now consistently pass and not depend on watching the
flush directory succeeding before the watcher is converted into a
stream.
Did some manual testing to make sure that we don't exit if glob watching
fails and instead report an error to clients:
From the daemon logs:
```
2023-05-23T15:36:56.894-0700 [TRACE] globwatch: watching "/private/tmp/turborepo-netflix-cache-repro"
2023-05-23T15:36:56.898-0700 [TRACE] globwatch: watching flush dir: "/private/var/folders/3m/rxkycvgs5jgfvs0k9xcgp6
km0000gn/T/turbod/3e1692e405755b5a/flush"
ERROR faked failure for globawatch
ERROR Globwatch config error: ServerFailedToStart
2023-05-23T15:37:21.582-0700 [TRACE] log: registering event source with poller: token=Token(1), interests=READABLE
| WRITABLE
```
From the turbo output:
```
web:build: Failed to check if we can skip restoring outputs for web#build: rpc error: code = Internal desc = failed to watch flush directory. Proceeding to check cache
```
---------
Co-authored-by: Chris Olszewski <Chris Olszewski>
