Data time subsetting

[SarahAlidoost]

closes #62
closes #42

🔴 this branch is based on #54, that should be merged first.

In this PR:

• monthly data are integrated to batch dimension
• the bias/scale (per-month parameters) are removed to make the model more flexible. We are using residuals training, so the mean seasonal signal is already removed. We are also injecting strong seasonal structure inside the network via cyclic time embedings, temporal attention over days, cross-month mixing in the aggregator.
  So the model is already learning a seasonal climatology and modulation, not relying on a fixed per-month correction.

[meiertgrootes]

See comment in PR #54 on replacing month positional encoding with a month-of-year feature in cyclical time embedding