Nightly binaries

[cpuhrsch]

In general the target plaform is restricted to linux (Ubuntu) and CUDA 10.2. We might want to ship a cpu-only package for faster testing and smaller binaries which is useful in particular for online tutorials or ease of installation.

Since this is built on top of PyTorch as a submodule, it'll ship with its own version of PyTorch. That means the user cannot install this package and also the regular PyTorch package.

• Ship pip wheel built from PyTorch submodule
• Ship conda wheel built from PyTorch submodule (nice to have)
• Add CI job(s) to run PyTorch submodule unittests on CPU
• Add CI job(s) to run PyTorch submodule unittests on CUDA
• Add CI job for nestedtensor unittests based on CPU submodule
• Add CI job for nestedtensor unittests based on CUDA submodule
• Conflict in some easily detectable way with an existing installation of PyTorch
• Many users install pytorch via conda: can we cause a conflict if this is the case?
• Define a minimum config for the pip package with environment flag (+nested)
• Create nightly s3 uploads of circleci binary artifacts
• Set special PyTorch version for submodule fork (e.g. 1.7.0-nestedtensor) and check in the package for it

Related PRs

• Add PyTorch fork with NestedTensor support as submodule + softmax via _foreach_softmax #239
• Add PyTorch test_nn and test_torch #263
• Save wheels as test job artifacts #265
• [DO NOT MERGE] Getting signal for nestedtensor fork  pytorch#47230