Summary:

This PR adds initial support for the 8-bit floating point types, matching
the spec in https://arxiv.org/pdf/2209.05433.pdf . We are adding this to
PyTorch to enable easier experimentation with these dtypes by the research
community.

Note: we do not yet have an RFC on a full end to end experience in PyTorch
for training and inference with float8, as it is too early to know what a
good UEX will look like. Specifically, we need to better understand the
performance and accuracy implications of scaling factor calculation to inform
the UEX design.

Note: we are aware that there are other specifications of float8, and
float8 is not yet in the IEEE spec. We are adding these dtypes to facilitate
easier experimentation with the hardware which is currently available, and
we are happy to follow-up on other float8 flavors when there is a need.

Note: the scaling factor is outside of the `float8` dtypes, and is expected
to be implemented by the user workflow.

Note: the current PR only has `torch.float8_e4m3fn`, to speed up initial review
and implementation of any requested changes. Once this passes initial review,
I will add `torch.float8_e5m2` in the same manner to this PR. It will be easiest
to land the two dtypes together due to some Meta-only changes which will
have to be done on the Meta-only version of this PR.

The specific support added here is:

```
// tensor creation with `torch.zeros`
x = torch.zeros(4, dtype=torch.float8_e4m3fn)

// conversion to and from float8 dtypes
// conversion between float32 and float8 uses the kernels copied from fbgemm
// conversion between other dtypes and float8 converts to float32 as an intermediate step
x = torch.randn(4, dtype=torch.float)
x_float8 = x.to(torch.float8_e4m3fn)
x_float32 = x_float8.to(torch.float32)

// printing out the tensor for debugging
// note: the values are converted to float32 before being printed
print(x_float8)

// numerical correctness of the format
TODO document
```

The following is not in this PR but which we plan to address in the near
future:
* ensure all of the functionality above works on CUDA

Test plan:

```
python test/test_quantization.py TestFloat8
```