Bitpacking default pack and unpack dimensions do not round-trip multidimensional tensors #4815

Bug

The optimized pack() and unpack() helpers have incompatible default dimensions: pack(..., dim=-1) packs the last dimension, while unpack(..., dim=0) expands the first dimension. As a result, the documented default round trip only works for one-dimensional tensors.

Reproduction

import torch
from torchao.prototype.dtypes.uintx.bitpacking import pack, unpack

x = torch.randint(0, 16, (8, 16), dtype=torch.uint8)
packed = pack(x, 4)
restored = unpack(packed, 4)

print(x.shape, restored.shape)

On current main (36027d82d), this prints:

torch.Size([8, 16]) torch.Size([16, 8])

The CPU reference pair (pack_cpu / unpack_cpu) already uses dim=-1 consistently, and TorchAO's internal UIntxTensor callers pass an explicit dimension.

Proposed fix

Change unpack()'s default to dim=-1 so the optimized pair has the same default contract as the CPU pair, and add a multidimensional round-trip regression covering bit widths 1 through 7. I have this small fix ready.