Since the _check_tfr_param function changes some of the values it
checks, scikit-learn estimators must perform this check at runtime, not
during estimator instantiation. Failing to do so leads to an error.

If the constructor modifies non-default parameters, scikit-learn will
throw an error during cloning.

Checking and transforming in the constructor violates sklearn
best-practice. Since this check is already performed in the
`_compute_tfr` method, we can simply remove the check from the
transformer class.

Since options checking doesn't modify the value it makes sense to keep
it in the constructor.

…n into time_frequency