Avoid Dashmap in favor of a shared threadsafe tree implementation

[Byron]

Most operations we do are actually incrementing counters. This should be as fast as possible, which is Arc<AtomicUsize>.

Other operations can be slower as they happen much more rarely.

In order to achieve this, DashMap should be replaced with a shared mutable tree implementation for prodash::tree, which is Send + Clone (but doesn't have to be sync). If the new implementation happens to be sync, which it could be, we can consider making this a property of prodash, which previously was a bit inconvenient as it forced users sometimes to wrap tree items into a mutex themselves.

Doing this naturally affects a lot of code especially in the renderers and it would be wise to figure out which steps to take to make this more gradual.

Benchmark Results

When using an Arc<AtomicUsize> compared to the current dashmap based implementation.

                        time:   [34.193 ns 34.199 ns 34.206 ns]
                        thrpt:  [146.17 Melem/s 146.20 Melem/s 146.23 Melem/s]
                 change:
                        time:   [-90.711% -90.674% -90.626%] (p = 0.00 < 0.05)
                        thrpt:  [+966.81% +972.27% +976.50%]
                        Performance has improved.
Found 13 outliers among 100 measurements (13.00%)
  2 (2.00%) low severe
  3 (3.00%) high mild
  8 (8.00%) high severe

[Byron]

Released in v16.1