Handle empty tensors (must bind at minimum 4 bytes on wgpu)

[laggui]

Some operations could return empty tensors (e.g., argwhere for a tensor where all elements are zero).

Trying to return an empty tensor with 0 byte alloc would fail on wgpu, as encountered in Burn issues (2180, 1946).

We can handle empty tensors by allocating 4 bytes (minimum required by wgpu).

We get the same result as other backends:

Tensor {
  data:
[],
  shape:  [0, 1],
  device:  BestAvailable,
  backend:  "fusion<jit<wgpu>>",
  kind:  "Int",
  dtype:  "i32",
}

[ArthurBrussee]

Wouldn't this also be an issue for kernels using client.empty()? Eg. if after no_true_vals.argwhere() I call .float().exp().

I thought the memory allocator handled allocating buffers of some minimum size behind the scenes.