Merge wrapper and implementations

[ev-br]

Undo the split between decorated functions and torch implementations: the formers are nothing but fall-throught to tha latters now that all argument pre/postprocessing is contained in normalization decorators.
While at it, merge _funcs.py and _wrapper.py, the split was purely technical.
The diff looks scary, but this is mainly just shuffling things around and trimming a small amount of dead weight.

[lezcano]

The merge LGTM. We are still missing to merge the other files within _detail, but this first step looks good.

I took this chance to go through the implementations here and flag a few issues with them. Just minor things all throughout.

Style Nit. Very often you name a variable result, just to return it in the next line. I think it's cleaner and more concise to simply do return torch.fn.

There's a number of places where we use methods rather than their torch.fn versions. It'd be good to be consistent all throughout. I flagged a few of these, but there's more (roll, squeeze, reshape...)

[lezcano]

What does this construction mean?

[ev-br]

if a is False: return False?

[lezcano]

Yeah, but a is an ArrrayLike, when is it going to be the literal False?

[ev-br]

OK, fair enough. Thanks for spotting this. Removed.

[lezcano]

This uncovered a latent bug:

    @normalizer(return_on_failure=False)
    def isscalar(a: ArrayLike):
        # XXX: this is a stub
>       return a.numel() == 1
E       AttributeError: 'bool' object has no attribute 'numel'

It may be better to just implement this function without the normalizer and remove the return_on_failure flag, as you initially had?

[ev-br]

All right. Here's what's going on: isscalar(np.int) (np.intis not a scalar!) fails to convert its argument to tensor, and the return_on_failure code path feeds False into the the body of the function. This is in fact requested by the normalizer argument three lines above, via normalizer(return_on_failure=False). So, can either

• put back the if a is False: return False clause + a code comment to not confuse future readers
• rip the return_on_failure code path and normalize manually (and possibly will need to normalize manually other functions from the not-yet-implemented numpy apis, who knows)

Your preference @lezcano ?

[lezcano]

Probably the simplest option is the second one, as this function can be implemented as

def isscalar(a):
    try:
        return asarray(a).size == 1
    except:
        return False

and it's very much one of a kind.

[ev-br]

OK, remove_on_failure is gone.

[lezcano]

Missing two minor things. Approving this contingent on fixing those two.

This one got us -600LOC, not too bad!

[ev-br]

CI is green, the two last comments are addressed, merging.