bug for stream-level metrics in periodic eval

[AntonioCarta]

periodic eval is broken in some subtle cases. Instead of calling eval on the entire stream, it calls it multiple times for each experience, breaking stream-level metrics.

Reproduce:

from avalanche.benchmarks import SplitMNIST
from avalanche.training import Naive
from avalanche.models import SimpleMLP

if __name__ == '__main__':
    benchmark = SplitMNIST(5)
    model = SimpleMLP()
    strat = Naive(model, None, eval_mb_size=512, eval_every=1)

    stream = benchmark.test_stream
    strat.train(stream[0], eval_streams=list(stream[:2]))

outputs:

-- >> Start of training phase << --
-- >> Start of eval phase << --
-- Starting eval on experience 0 (Task 0) from test stream --
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:05<00:00,  1.08s/it]
> Eval on experience 0 (Task 0) from test stream ended.
        Loss_Exp/eval_phase/test_stream/Task000/Exp000 = 2.3881
        Top1_Acc_Exp/eval_phase/test_stream/Task000/Exp000 = 0.0033
-- >> End of eval phase << --
        Loss_Stream/eval_phase/test_stream/Task000 = 2.3881
        Top1_Acc_Stream/eval_phase/test_stream/Task000 = 0.0033
-- >> Start of eval phase << --
-- Starting eval on experience 1 (Task 0) from test stream --
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:01<00:00,  2.12it/s]
> Eval on experience 1 (Task 0) from test stream ended.
        Loss_Exp/eval_phase/test_stream/Task000/Exp001 = 2.3041
        Top1_Acc_Exp/eval_phase/test_stream/Task000/Exp001 = 0.2129
-- >> End of eval phase << --
        Loss_Stream/eval_phase/test_stream/Task000 = 2.3041
        Top1_Acc_Stream/eval_phase/test_stream/Task000 = 0.2129

which means it's calling eval two times instead of a single one with the entire stream.