Incorrect task_n_future_covariates Counting and Misordered Covariate Rows in validate_and_prepare_single_dict_task

[HarvestStars]

Describe the bug

There are two logic bugs in the function validate_and_prepare_single_dict_task, which can lead to incorrect covariate alignment and wrong task_n_future_covariates counting.

1. Incorrect task_n_future_covariates counting
   The function currently sets:

   task_n_future_covariates = len(task_future_covariates_list)

   However, task_future_covariates_list is built by iterating through all past_covariates keys (both past-only and known-future).
   This makes task_n_future_covariates incorrectly equal to the total number of covariates, not the number of known-future covariates.
   The returned value thus disrupts the following function _construct_slice, and its task_n_past_only_covariates, which are for bacth-building when trainning.

2. Covariate row order not guaranteed
   Even though the keys are sorted, the current logic does not ensure that “known-future” covariates appear as the last rows in task_future_covariates_tensor. Also mislead _construct_slice in

       task_future_covariates = task_past_tensor[
             -task_n_future_covariates:, slice_idx : slice_idx + self.prediction_length
         ]

---

Expected behavior

• task_n_future_covariates should reflect only the number of known-future covariates (len(task_future_covariates_keys)).
• Both tensors (task_context_tensor, task_future_covariates_tensor) should have consistent row ordering:
  • Past-only covariates first
  • Known-future covariates last

---

To reproduce

You can reproduce the issue using the following minimal example:

task = {
    "target": np.arange(10),
    "past_covariates": {
        "temp": np.arange(10),
        "holiday": np.array(["yes", "no", "no", "yes", "no", "yes", "no", "no", "no", "yes"]),
    },
    "future_covariates": {
        "holiday": np.array(["no", "yes", "yes", "no"]),
    },
}

out = validate_and_prepare_single_dict_task(task, idx=0, prediction_length=4)
print(out[-1])  # task_n_future_covariates

Expected result: 1 (only holiday has future values)
Current result: 2 (counts both temp and holiday)

---

Proposed fix

This PR(#344) separates past-only and future-known covariates explicitly, and fixes the counting logic:

Code diff

# Separate keys to ensure correct ordering
past_only_keys = [k for k in task_past_covariates_keys_all if k not in task_future_covariates_keys]
ordered_covariate_keys = past_only_keys + task_future_covariates_keys

# Build tensors in the same order
for key in ordered_covariate_keys:
    ...

# Correct counting logic
task_n_future_covariates = len(task_future_covariates_keys)