Fix confusion_matrix raising on tensor sample weights

[RicardoSantos-99]

Scholar.Metrics.Classification.confusion_matrix/3 takes :sample_weights as a list but raises on a tensor:

y_true = Nx.tensor([0, 0, 1, 1], type: :u32)
y_pred = Nx.tensor([0, 1, 1, 1], type: :u32)

Classification.confusion_matrix(y_true, y_pred, num_classes: 2, sample_weights: [1, 2, 1, 1])
#Nx.Tensor<
  s64[2][2]
  [
    [1, 2],
    [0, 2]
  ]
>

Classification.confusion_matrix(y_true, y_pred, num_classes: 2, sample_weights: Nx.tensor([1, 2, 1, 1]))
** (ArgumentError) invalid numerical type: nil (see Nx.Type docs for all supported types)

A scalar weight tensor fails the same way. Both shapes are accepted input: Scholar.Options.weights validates rank 0 and rank 1 tensors, and validate_weights/3 lists [{}, {num_samples}] among the shapes it handles. So the option accepts shapes the implementation then rejects, and the error names none of it.

Shared.validate_weights/3 reads :type from its opts, which is optional. The nil and list branches work without it, since Nx.tensor(weights, type: nil) infers the type. The tensor branch calls Nx.as_type(weights, nil), which raises. Of the ten call sites, confusion_matrix is the only one that passes no type, so it is the only one that reaches this.

The fix converts only when a type is given, which leaves the tensor branch behaving like the list branch: the weights keep their own type.

Fixing it at the call site instead, passing to_float_type(y_true) like the other callers do, also works but changes what the function returns. An unnormalized count matrix built from integer weights would come back as floats.

Two tests cover it: tensor weights matching the same weights as a list, and a scalar weight applying to every sample.

[josevalim]

💚 💙 💜 💛 ❤️