Description

Two small, related fixes to pypots/nn/functional/error.py that make
calc_quantile_loss behave like its siblings in the same module.

1. Missing _check_inputs call.
Every other calc_* in the file (calc_mae, calc_mse, calc_rmse,
calc_mre, calc_quantile_crps, calc_quantile_crps_sum) starts with
lib = _check_inputs(...) so NaN / shape / dtype problems surface as a
clear AssertionError. calc_quantile_loss was the one exception and
silently returned tensor(nan) on NaN input, e.g.

calc_mae(p_with_nan, t, m)              # AssertionError, good
calc_quantile_loss(p_with_nan, t, 0.5, m)  # tensor(nan), bad

Adding the same guard makes the module's validation contract uniform.

2. numpy-in / numpy-out parity.
The function signature returns Union[float, torch.Tensor], matching the
sibling metrics which preserve the caller's type (numpy in → numpy out).
After #822 added numpy support, calc_quantile_loss always converted
numpy inputs to torch tensors and returned a torch.Tensor, breaking
that contract:

type(calc_mae(np_a, np_b))             # numpy.float32
type(calc_quantile_loss(np_a, np_b, 0.5, np_m))  # torch.Tensor  <-- regression

Restoring the numpy return on numpy inputs keeps downstream code that
expected a plain numpy scalar working.

Changes

• pypots/nn/functional/error.py — calc_quantile_loss:
  • call _check_inputs(predictions, targets, eval_points) first;
  • capture numpy_in = isinstance(predictions, np.ndarray) before the
    conversion block;
  • when numpy_in, return quantile_loss.detach().cpu().numpy() so the
    declared Union[float, torch.Tensor] contract holds.

The change is confined to one function and does not touch the
_check_inputs signature, so it does not conflict with the in-flight #821.

Testing

• NaN input now raises AssertionError: predictions mustn't contain NaN values, matching calc_mae / calc_mse / …
• Shape mismatch raises AssertionError: shape of predictions and targets must match … instead of relying on downstream broadcast errors.
• numpy in → numpy out is preserved; torch in → torch out is unchanged.
• calc_quantile_crps still calls calc_quantile_loss internally on
  torch tensors; that path is numerically unchanged.