Dask-CuDF: use default Dask Dataframe optimizer (#8581)

In order to use the new HighLevelGraph optimization work in Dask/Distributed, this PR makes `dask_cudf.Dataframes` use the default Dask optimizer.
Previously, we have been explicitly materialized the HighLevelGraphs when calling `submit()` and `compute()` on `dask_cudf.Dataframes`.

Overall, this should improve performance but by default low-level task optimizations are disabled, which _might_ have a negative impact. High-level optimizations are done in any case and we are working on moving all low-level optimization to high-level but currently low-level optimization such as array slicing is only supported by the low-level. 

I don't think we will be missing any low-level optimizations related to Dataframes so I think we should follow Dask on this one and disable low-level optimizations by default.
It is possible to enable low-level optimizations explicitly by setting the Dask config like:
```python
dask.config.set({"optimization.fuse.active": True})
```

cc. @jakirkham, @quasiben, @beckernick, @VibhuJawa

Authors:
  - Mads R. B. Kristensen (https://github.com/madsbk)

Approvers:
  - https://github.com/jakirkham

URL: #8581

python/dask_cudf/dask_cudf/core.py

@@ -9,10 +9,9 @@
 from tlz import partition_all
 
 import dask
+import dask.dataframe.optimize
 from dask import dataframe as dd
 from dask.base import normalize_token, tokenize
-from dask.context import _globals
-from dask.core import flatten
 from dask.dataframe.core import (
     Scalar,
     finalize,
@@ -22,7 +21,6 @@
 )
 from dask.dataframe.utils import raise_on_meta_error
 from dask.highlevelgraph import HighLevelGraph
-from dask.optimization import cull, fuse
 from dask.utils import M, OperatorMethodMixin, apply, derived_from, funcname
 
 import cudf
@@ -34,19 +32,6 @@
 DASK_VERSION = LooseVersion(dask.__version__)
 
 
-def optimize(dsk, keys, **kwargs):
-    flatkeys = list(flatten(keys)) if isinstance(keys, list) else [keys]
-    dsk, dependencies = cull(dsk, flatkeys)
-    dsk, dependencies = fuse(
-        dsk,
-        keys,
-        dependencies=dependencies,
-        ave_width=_globals.get("fuse_ave_width", 1),
-    )
-    dsk, _ = cull(dsk, keys)
-    return dsk
-
-
 class _Frame(dd.core._Frame, OperatorMethodMixin):
     """ Superclass for DataFrame and Series
 
@@ -65,7 +50,6 @@ class _Frame(dd.core._Frame, OperatorMethodMixin):
     """
 
     __dask_scheduler__ = staticmethod(dask.get)
-    __dask_optimize__ = staticmethod(optimize)
 
     def __dask_postcompute__(self):
         return finalize, ()