Firecracker VFS fixes / improvements

[bduffany]

Polish up VFS a bit so that we can experiment with using VFS instead of workspace ext4 images. This should help reduce the IO costs associated with mke2fs, but at the cost of higher file access latency during execution.

• Add test
• Improve error handling when guest VFS server fails:
  • Switch log.Error to log.Fatal in the init process when the VFS server crashes. (We currently only show fatal-level logs.)
  • If VFS setup in the guest fails, don't hang forever in SendPrepareFilesystemRequest. Instead, parse the fatal error from the VM logs and report it immediately. This fix involved some cleanup/refactoring to the VM startup code.
• Remove unused output fields from FileSystemLayout. This struct was missing the new output_paths field, but since the output_files and output_directories aren't used anyway, we can just remove them.
• Fix a few cases where we don't respect whether VFS is enabled or not:
  • If VFS is enabled for the execution, then don't mount the workspace drive at all, even the "stub" one mounted initially.
  • If VFS is enabled by the executor but not enabled via platform props, then don't unnecessarily start the VFS server in the guest.