Adds minimal support for sparse half embedding on CPU (backward() and sum()). Previous PR only allowed sum().backward() of sparse half embedding on CUDA. We can add more operators if this is useful, but currently skips most CPU operator implementations that depend on blas or vec256 since they don't have float16 support.

After 19695,

>>> a = torch.nn.Embedding(3, 4, sparse=True).half()
>>> a(torch.LongTensor([1, 0])).backward(torch.ones(2,4).half())

Gave: RuntimeError: _th_index_select not supported on CPUType for Half
and a(torch.LongTensor([1, 0]).sum() gave an error that sum_cpu was not available.

Builds on: #19695
Don't need to re-review commits before 29a72ea