[FEA] Improve stopping condition for SVM training

[tfeher]

Is your feature request related to a problem? Please describe.
In certain cases the SVM training fails to converge to the target tolerance in a reasonable time.
One such case is the following:

iris = load_iris()
X = iris.data
y = iris.target
# we only keep the first two  features in X and sub-sample the dataset to keep 
# only 2 classes and make it a binary classification problem.
X_2d = X[:, :2]
X_2d = X_2d[y > 0]
y_2d = y[y > 0]
y_2d -= 1
scaler = StandardScaler()
X = scaler.fit_transform(X)
X_2d = scaler.fit_transform(X_2d)
params = {'kernel':'poly', 'C': 10, 'gamma': 1e1, 'verbose':1}
sklSVC = sklearn.svm.SVC(**params)
sklSVC.fit(X_2d, y_2d)
cuSVC = cuml.svm.SVC(**params)
cuSVC.fit(X_2d, y_2d)

Likely cause is an oscillation in the diff parameter. Such oscillation could happen if the same set of vectors are getting in and out of the working set,

Describe the solution you'd like

• Detect if there is an oscillation in the diff parameter
• Stop training in such case

Describe alternatives you've considered
Currently the training would stop if the number of maximum iterations are reached.

Additional context
This is slightly related to #946, because the working set selection method influences what kind of problems can occur with the convergence.

[tfeher]

The training can still converge if the diff parameter oscillates. This will be addressed with a debug message which gives advice how to improve the convergence.