turn ncm to dynamic module that adapts with 0 vectors at evaluation

[AlbinSou]

Adds an adaptation at evaluation for NCM so that it fills class_means_dict with torch.zeros (default behavior for non-pre-existing classes).

Previously using eval_every=1 would fail unless all classes are present in the class_means_dict. Now it still fails for the first step (no classes in dict) since the feature size is unknown when no feature vectors are present. But it does not fail anymore for other steps.

[coveralls]

Pull Request Test Coverage Report for Build 5553432679

• 7 of 9 (77.78%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.3%) to 72.843%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/models/ncm_classifier.py	7	9	77.78%

Totals
Change from base Build 5478303948:	0.3%
Covered Lines:	16703
Relevant Lines:	22930

---

💛 - Coveralls