* [UnitTests] Require cached fixtures to be copy-able, with opt-in.
Previously, any class that doesn't raise a TypeError in copy.deepcopy
could be used as a return value in a @tvm.testing.fixture. This has
the possibility of incorrectly copying classes inherit the default
object.__reduce__ implementation. Therefore, only classes that
explicitly implement copy functionality (e.g. __deepcopy__ or
__getstate__/__setstate__), or that are explicitly listed in
tvm.testing._fixture_cache are allowed to be cached.
* [UnitTests] Added TestCachedFixtureIsCopy
Verifies that tvm.testing.fixture caching returns copy of object, not
the original object.
* [UnitTests] Correct parametrization of cudnn target.
Previous checks for enabled runtimes were based only on the target
kind. CuDNN is the same target kind as "cuda", and therefore needs
special handling.
* Change test on uncacheable to check for explicit TypeError
Lunderberg
dbf9ce52d4