[NestedTensor] multiply batch and ragged dimension to get shape of values tensor

[davidberard98]

🚀 The feature, motivation and pitch

From @albanD:

I would argue that it would be so cool to be able to do something like:

# Assume I want a new Tensor with the first two dimensions collapsed
nested_size = my_nested_tensor.size()
nested_size[1] = nested_size[0] * nested_size[1]
nested_size.pop(0)
my_new_nested_tensor = torch.zeros(nested_size)

#124687 (comment)

Alternatives

I think this is just the same as nt.values().

Additional context

No response

cc @cpuhrsch @jbschlosser @bhosmer @drisspg @soulitzer

[albanD]

cc @ezyang because I guess that's in line with what you meant during Triage meeting saying that NestedInt should be be a first class citizen?

[ezyang]

Not really lol.

This is an interesting request. I'm not sure it can actually be doable this way though. There are some equations that don't hold in @bhosmer style fold formulation. With fold semantics, (B, *, E) is represented as (B, [s1, s2, s3, ... sb], E), but we already want scalar multiplication by nested int to denote scaling the nested int by that amount (because this is good for stride formulas).