Torchvision ops registration still not fixed in C++

[bmanga]

It turns out the solution I implemented for automatic ops registration in C++ is not robust enough.
If falls apart when the actual torchvision code is enclosed in a static library, which is then meant to be consumed by an executable. This indirect hop through a library exposes the pruning problem all over again. Apologies for not realizing earlier and implementing it this way :/.

In the end, I think the only two viable solutions are the ones we discussed a while ago:

1. Force include all symbols through linker commands (eg. --whole-archive)
2. Add, only for the C++ API, a register_operators function that must be called in order to explicitly load the operators, and stop relying on the implicit behavior

As before my preference goes to 2 because it's simpler to implement and maintain, and it doesn't risk bloating the binary with unneeded symbols. However, option 1 is what torch already does (sadly).

Thoughts? @fmassa @datumbox

[fmassa]

I would say to go for option 1 -- we are doing this internally already, and once (if) pytorch changes the behavior to something a bit more optimal we can follow the same approach.

[datumbox]

What side-effects (if any) do we have if we go with option 1 for mobile devices where the size of the binary is important? Will this conflict with other initiatives that try to reduce the number of registered ops on mobile?

[fmassa]

@datumbox for now mobile has its own CMakeLists, so this change wouldn't affect mobile directly