Fix gather_tensor_shape discarding zero-sized dimensions and crash in copy_tensor_to_devices

[Ryanakml]

Problem

1. gather_tensor_shape silently corrupts shapes for tensors with zero-sized dimensions (e.g. empty sequence slices or empty batch tensors):
   Passing t = torch.zeros(3, 0, 4) into gather_tensor_shape(t) returns shape [3, 4] instead of (3, 0, 4) — the zero dimension is completely stripped.
2. copy_tensor_to_devices crashes with TypeError on worker ranks:

  File "accelerate/utils/operations.py", line 600, in copy_tensor_to_devices
    tensor = torch.zeros(shape, dtype=TENSOR_INT_TO_DTYPE[dtype]).to(state.device)
TypeError: zeros(): argument 'size' (position 1) must be tuple of ints, not Tensor

Root cause

1. gather_tensor_shape used base_tensor[base_tensor.nonzero()] to filter non-zero elements. Any legitimate 0 dimension (such as (0, 10) or (3, 0, 4)) gets stripped out along with padding.
2. In PyTorch, indexing a 1D tensor with its .nonzero() indices returns a 2D tensor of shape (N, 1). When worker ranks (tensor is None) call copy_tensor_to_devices, torch.zeros(shape, ...) fails because shape is a 2D tensor rather than a tuple of ints.
3. torch.empty was previously allocated for base_tensor, which risks reading uninitialized dirty memory from the PyTorch caching allocator on ranks where tensor is None.

Fix

• Explicitly encode num_dims + 1 at index 0 of base_tensor and write/read shape dimensions directly by slicing without relying on nonzero().
• Allocate safely with torch.zeros (sized to 66 elements, since standard PyTorch supports up to 64 dimensions).
• Return shape as a standard tuple of ints, ensuring torch.zeros(shape, ...) works properly across all worker ranks.

Verification

• Added test_gather_tensor_shape and test_copy_tensor_to_devices in tests/test_utils.py testing scalars (), multi-dim tensors, and tensors with zero-sized dimensions ((3, 0, 4) and (0, 10)).
• All 50 tests in tests/test_utils.py pass (50 passed, 4 skipped).
• Code formatted and clean with ruff format --check.