Set GPU ThreadPoolExecutor and set known libraries to use it

With @madsbk 's recent work allowing for multiple executors, we might consider making a GPU `ThreadPoolExecutor` in workers by default if a GPU is detected, and then annotating tasks known to be GPU-targetted with that annotation. This would improve the likelihood that a user of vanilla dask has a good time with RAPIDS, cupy, or other known project.

We probably can't do this in full generality (it's hard to detect what code has GPUs) but we're no worse off if we don't catch something, and we can handle the common cases well.

Concretely, I propose:

1. Having the Worker class try importing pynvml (or some future NVIDIA python library) and if it detects a GPU then create a single-threaded ThreadPoolExecutor

```python
try:
import pynvml
except ImportError:
pass
else:
if pynvml.do_I_have_a_gpu():
self.executors["gpu"] = ThreadPoolExecutor
```

2. In known GPU libraries we would add an annotation to every layer

```python
class ArrayLayer:
def __init__(self, ...):
if "cupy" in str(type(self._meta)):
self.annotations["executor"] = "gpu" # or use setdefault or something
```

cc @madsbk @quasiben @kkraus14

dask-cuda handles this for users who use it. This feels like something that we could upstream. This would also help with CPU-GPU mixed computation.