Host-wide measures are double-counted in the GUI

There are some problems with totalizations in the Bokeh GUI whenever there's more than one worker per host. They are particularly glaring on LocalClusters.

- In the `Workers` tab:
- The cluster total for the columns `net read`, `net write`, `disk read` and `disk write` sums up the value for each worker. However, these are host-wide measures, so if two or more workers sit on the same host, the total will be double-counted.
- The cluster total for the columns `gpu_memory_used` and `gpu_utilization` sums up the value for each worker. However, two workers may share the same GPU, depending on the `CUDA_VISIBLE_DEVICES` environment variable (see `nvml.py`). If the variable is not set and in general on single-GPU hosts, all workers on the same host will share the same GPU. Again, this leads to double-counting.
- The cluster total of the column `event_loop_interval` is a sum of the workers. This makes no sense; it should be a mean.
- The `More... -> Workers Disk` and `More... -> Workers Network` tabs show one bar per worker. This is misleading; there should be one bar per host.
- The `More... -> GPU Memory` and `More... -> GPU Utilization` tabs show one bar per worker. This is misleading; there should be one bar per GPU.