Out of sample testing

[ArneTR]

This PR builds upon the bin approach in #4

It adds the --test-data param that points to a testing only energy logger data file.

This allows proper out of sample testing.

First results show this for non same workloads (idle vs benchmark.sh):

python3 model.py /tmp/energy-7hyUsZ9f.log --mode delta --alpha 10 --trim=1.0 --test-data /tmp/energy-3G05W1dt.log
# mode=delta alpha=10.0
# rows_train=1326 rows_test=418

# --- test metrics (current module defaults) ---
R2     : -15.25236
RMSE   : 946323.788 uJ
MAE    : 916790.084 uJ
MAPE   : 99.927 %

# --- test metrics (linear non-negative ridge) ---
R2     : -12.23006
RMSE   : 853812.158 uJ
MAE    : 830840.973 uJ
MAPE   : 100.021 %

# --- test metrics (nonlinear reference) ---
R2     : -0.43362
RMSE   : 281059.228 uJ
MAE    : 254916.150 uJ
MAPE   : 29.056 %

# --- test metrics (distilled kernel model) ---
R2     : -31.53800
RMSE   : 1338988.676 uJ
MAE    : 795537.625 uJ
MAPE   : 94.943 %

# --- train metrics (distilled kernel model) ---
R2     : -575.81784
RMSE   : 44334482.657 uJ
MAE    : 2338971.645 uJ
MAPE   : 79.459 %

For repeated runs of benchmark.sh it is this:

arne@framebook:~/Sites/procpower/tools$ python3 model.py /tmp/energy-7hyUsZ9f.log --mode delta --alpha 10 --trim=1.0 --test-data /tmp/energy-gKpkxu81.log
# mode=delta alpha=10.0
# rows_train=1326 rows_test=1246

# --- test metrics (current module defaults) ---
R2     : -2.88214
RMSE   : 3388222.120 uJ
MAE    : 2922949.824 uJ
MAPE   : 99.826 %

# --- test metrics (linear non-negative ridge) ---
R2     : 0.86565
RMSE   : 630306.004 uJ
MAE    : 576792.573 uJ
MAPE   : 32.694 %

# --- test metrics (nonlinear reference) ---
R2     : 0.96917
RMSE   : 301920.783 uJ
MAE    : 261693.641 uJ
MAPE   : 12.957 %

# --- test metrics (distilled kernel model) ---
R2     : 0.20774
RMSE   : 1530631.787 uJ
MAE    : 625044.750 uJ
MAPE   : 33.205 %

# --- train metrics (distilled kernel model) ---
R2     : -575.81784
RMSE   : 44334482.657 uJ
MAE    : 2338971.645 uJ
MAPE   : 79.459 %