[ENH] Refactor N-BEATS blocks to separate KAN logic (#2012)

#### Reference Issues/PRs
Fixes #2011

---
#### What does this implement/fix? Explain your changes.
This PR refactors the N-BEATS architecture to explicitly separate
standard MLP blocks from KAN-based blocks, removing the previous
`if/else` dependency on the `use_kan` flag.

* **Split Classes:** Created distinct classes for all variants to ensure
clean separation of concerns:
    * `NBEATSBlock` / `NBEATSBlockKAN`
    * `NBEATSSeasonalBlock` / `NBEATSSeasonalBlockKAN`
    * `NBEATSTrendBlock` / `NBEATSTrendBlockKAN`
* **Mixin Pattern:** Implemented `SeasonalMixin` and `TrendMixin` to
handle the shared mathematical logic (Fourier series and Polynomial
generation) between standard and KAN blocks.
* Unpacked the `kan_params` dictionary into explicit `__init__`
arguments (e.g., `num`, `k`, `noise_scale`) in the KAN classes.
* Updated `NBeatsKAN` to correctly initialize these new specific block
classes based on the `stack_types` argument.
---
#### What should a reviewer concentrate their feedback on?
* Please verify that `SeasonalMixin` and `TrendMixin` correctly capture
the shared logic without unwanted side effects.
* Check if the use of `**kan_kwargs` in the subclasses correctly passes
parameters up to `NBEATSBlockKAN`.
* I restored `F.relu` in `NBEATSGenericBlockKAN` to ensure consistency
with the standard generic block.
---
#### Did you add any tests for the change?
Yes, I added `test_nbeats_kan_integration` which covers the following:
1. Verifies Init -> Train -> Save -> Load -> Predict flows working
together.
2. Asserts that KAN-specific hyperparameters (e.g., `num=10`) are
correctly saved to `hparams` and restored upon checkpoint loading.
3. Asserts that the model actually initializes the correct class types
(e.g., `NBEATSTrendBlockKAN`) when `stack_types` are requested.

Author: khenm

pytorch_forecasting/layers/_nbeats/__init__.py

@@ -4,14 +4,22 @@
 
 from pytorch_forecasting.layers._nbeats._blocks import (
     NBEATSBlock,
+    NBEATSBlockKAN,
     NBEATSGenericBlock,
+    NBEATSGenericBlockKAN,
     NBEATSSeasonalBlock,
+    NBEATSSeasonalBlockKAN,
     NBEATSTrendBlock,
+    NBEATSTrendBlockKAN,
 )
 
 __all__ = [
     "NBEATSBlock",
     "NBEATSGenericBlock",
     "NBEATSSeasonalBlock",
     "NBEATSTrendBlock",
+    "NBEATSBlockKAN",
+    "NBEATSGenericBlockKAN",
+    "NBEATSSeasonalBlockKAN",
+    "NBEATSTrendBlockKAN",
 ]