Allow to write group-level lag updates without write-locking the whole LagRegister

[detro]

Rework the LagRegister internal data structure to hold a RwLock on the GroupWithLag level.

Currently the LagRegister looks like this:

pub struct LagRegister {
    pub(crate) lag_by_group: Arc<RwLock<HashMap<String, GroupWithLag>>>,
}

pub struct GroupWithLag {
    pub(crate) group: Group,
    pub(crate) lag_by_topic_partition: HashMap<TopicPartition, LagWithOwner>,
}

When we need to update lag information for a specific group, we are holding a whole write-lock on the LagRegister:

let mut w_guard = lag_register_groups.write().await;

We need to rework this so that each GroupWithLag is stored in the lag_by_group map, wrapped into its own RwLock. This way we could access the LagRegister first-level map with a read-lock, and only request the .write().await lock on the GroupWithLag.

This would enable concurrent updates of the register, enhancing performance and reducing lock contention drammatically.