Add cuda tests

[ValerianRey]

This PR adds the possibility to run unit tests on cuda (fixes #97).

• 7284c0b2aa850174ef6cf0703f0da9a5415d3484 Add DEVICE constant in conftest.py
• e505a393ebc3335eb7820d6321512b051ed3f364 Change unit tests to move tensors to DEVICE
• 40701b2c1066d68846c3529cd45e61dd938b8888 Add command for cuda unit tests in CONTRIBUTING.md

As a result, a test that was already slow became even slower on cuda, so this PR also changes this test:

• 231cb004b8c427889945928521343fae6c214217 Make test_mgda_satisfies_kkt_conditions faster

Lastly, while developing this PR, I found several small issues (sometimes unrelated to the device):

• fe16544e88a142092df5428a5a1672c6fdad7986 Fix leftover print in test
• fd1e501a48ecb3e8d5f6d1fa1f55f1e68ff99bbd Fix wrong device and dtype in exceptional case in IMTL-G
• cf3a5ac2fcc93b522d63537a67364194f8601663 Fix wrong github link in error message
• 0ad72de5ae214e3f5ac244173ccc0bb12fb56321 Fix usage of legacy torch.Tensor constructor
• f68ada8b04be7c2fb70a94f09d80aae82826392a Fix wrong parameter name in _jac docstring
• da92580df016fbad4e723036bb76153eb4da2ab5 Fix not specifying the device and dtype when n_outputs==0 in _grad and _jac

In my opinion the only commit worth a changelog entry is fd1e501a48ecb3e8d5f6d1fa1f55f1e68ff99bbd, so for me the changelog is ready.