Refactor preprocessing logic for Chronos-2

[shchur]

Problem

Converting raw data to list[PreparedInput] is extremely slow when categorical covariates are present. For M5 (30K items, 50M rows, 7 categorical columns), preprocessing takes ~8 minutes. The bottlenecks are:

1. Per-item target encoding loop using sklearn
2. Redundant validation across validate_df_inputs and validate_and_prepare_single_dict_input

Proposed design

New module chronos/chronos2/preprocess.py with four entry points:

Entry point	Input format
from_tensor()	3D tensor
from_tensor_list()	list of 1D/2D tensors
from_dataframe()	pd.DataFrame + optional future_df
from_dict_list()	list[dict]

Key changes:

• Direct tensor conversion — no intermediate dict representation, no unnecessary validation
• Vectorized target encoding — _target_encode() uses bincount across all items at once instead of per-item sklearn calls
• Single validation point — each entry point validates once, internal methods assume valid input
• Removes sklearn dependency for encoding

Breaking changes

• Drops support for heterogeneous list[dict] inputs (all items must have same structure)

Expected speedup

~20x faster on M5 dataset (8 min → ~25s based on earlier prototyping)

---

This PR contains only the design (signatures + docstrings). Implementation to follow.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[shchur]

We could also introduce an optional kwarg here (e.g. future_covariates_names: list[str] | None) to specify the names of known covariates during training when future_df is unavailable.

Same for from_dict_list to remove the need for the {col: None for col in future_covariates_name} hack there.

[abdulfatir]

Thanks @shchur!

Overall, I don't have major concerns. The main thing we need to be fine with is that now there will be multiple validation paths which must be maintained. However, I think the major ones are list_of_dicts and dataframe, so it should not be too difficult to keep them tidy.

[abdulfatir]

Maybe list_of_dicts to align with the naming elsewhere?

[abdulfatir]

What's the difference between these two?

[abdulfatir]

What about NaNs in the categories (not in target)?