This PR fixes split_observations for conditional parameters to calculate EI values.

Motivation

The current implementation of the split_observations in the TPESampler is as follows:

1. The sampler discards the trials that do not contain the parameter named param_name when it gets observation pairs. This causes different len(config_vals) for conditional parameters.
   

   optuna/optuna/samplers/tpe/sampler.py

   Line 562 in f614d69

   if param_name not in trial.params:

2. Then, the sampler calculates the y* value based on the gamma = min(int(np.ceil(0.1 * len(config_vals))), 25)
   

   optuna/optuna/samplers/tpe/sampler.py

   Line 183 in f614d69

   n_below = self._gamma(len(config_vals))

3. Finally, the sampler splits config_vals into the ones with loss_vals < y* (i.e., the observations for l(x)) and the remaining ones (i.e., the observations for g(x)).

Based on my understanding, y* which determines the integration interval for the EI calculation should be the same value for all parameters (ref. the TPE paper).
However, the current strategy may set different y* values for conditional parameters.
Here is an example of this situation (note: this example is maximization).
In this example, y* for params_classifier is 0.9466666666666667, whereas y* for params_svc_c 0.32.
This seems to cause a curious EI calculation, i.e., a different integration interval is set for each parameter.

Description of the changes

This PR changes the split_observation procedure as follows:

1. The sampler does not discard the trials that do not contain the parameter named param_name when it gets observation.
2. Then, the sampler calculates the y* value based on the gamma = min(int(np.ceil(0.1 * len(config_vals))), 25). This y* is consistent among all hyperparameters.
3. After that, the sampler splits config_vals into the observations for l(x) and the observations for g(x).
4. Finally, the sampler discards the config_vals with the observations that do not contain the parameter named param_name from the observations for l(x) and g(x).

I have also confirmed that the split_observation procedure in the original TPE implementation in HyperOpt is like this.
https://github.com/hyperopt/hyperopt/blob/master/hyperopt/tpe.py#L625
In this implementation, o_vals is the set of config_vals that contain the parameter named param_name only, whereas l_vals is the set of all of the observed losses so far.

My apologies if I am misunderstanding the algorithm and please close this PR.