Use meta tensors for shape and dtype function tests (#1990)

There are some restrictions on the CPU device that prevent the current
dtype functions from working with graphs extracted using other
devices. For example, many ops don't support float16 on CPU, but they
do on CUDA. To avoid these restrictions, this commit switches the
abstract interpretation library testing framework to use the META
device. The META device is a lot more lenient in terms of what dtypes
are supported, runs computations on cpu, and because META tensors
don't carry any data, the actual checks are performed much
faster. This removes all the float16 restrictions.

I did notice some inconsistencies in the expected results for one op
when using CPU vs META, and some ops simply don't work on the META
device. I will open bugs upstream for the inconsistencies, and in a
follow up patch I will add a way for `@check`s to specify device to
test the ops that don't run on META. For now, I have commented out the
tests of those ops.

Inconsistent ops:
- aten.threshold_backward

Cannot run on META:
- aten._convolution
- aten._convolution.deprecated
- aten.bincount