Fix nested tensors predictions compatibility with fp16

[tcapelle]

Modifies retrieving first tensor to check float16 dtype to support nested tensors.

Warning: This PR needs the flatten function that is added in this other PR.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[tcapelle]

@jph00 do you know why the test is not passing? my bad, it is missing the fastcore integration of flatten PR

[jph00]

@tcapelle can you please try changing listify(flatten(...)) to concat(...), since that's the same thing now

[tcapelle]

@jph00 we have a conflict with torch_core.concat
I did a quick fix but found another bug, try:

retain_type([1,2,3], tuple([1]))
> [1,2,3]

it should be:

> (1,2,3)

[jph00]

OK happy to leave it as-is then. Well spotted on the concat name clash - that's actually been there for ages but no-one had noticed before!

[jph00]

Just another thought: what if instead of listify(flatten(self.pred))[0] we did next(flatten(self.pred)) ? That would avoid needlessly allocating extra memory.

[jph00]

Looks like it's all sorted! I'll update the fastcore version dep too.