Fix scalar broadcast bug in inductor lowering

[gmagogsfm]

Fixes a bug I discovered in vLLM Helion kernel authoring

InductorLowering was incorrectly expanding scalars (0-D tensors) with [None, None] to match the max ndim of all inputs. This created broadcast shape mismatches in generated Triton code like scale_val[None, None] when multiplying a 2D tensor by a scalar.

Fix: special case for 0-D tensors and expect Triton to handle scalar broadcasting

Example:

@helion.kernel(
    config=helion.Config(
        block_sizes=[2, 64],
        flatten_loops=[True],
        indexing=["pointer", "pointer", "tensor_descriptor"],
    )
)
def scalar_multiply(x: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
    m, n = x.shape
    out = torch.empty_like(x)
    for tile_idx in hl.tile(out.shape):
        scale_val = hl.load(scale, [0])  # Load scalar from 1-element tensor
        out[tile_idx] = x[tile_idx] * scale_val
    return out

result = scalar_multiply(torch.randn([4, 128]), torch.tensor([2.0]))

Before

  scale_val = tl.load(...)
  v_0 = scale_val[None, None]  
  result = x_tile * v_0 

  Error: Cannot broadcast, rank mismatch: ['4', '128'], ['1', '1']

After

  scale_val = tl.load(...)
  result = x_tile * scale_val  # ✅ Scalar broadcasts correctly to [4, 128]

[gmagogsfm]

@yf225 I think I hit a bug in Helion's handling of scalar tensor and attempted a fix. Could you check to see if this makes sense?

[yf225]

lgtm thanks! might need to fix lint

update: lint job seems broken on trunk, will try to see if I can repro and fix tomorrow