Add property-based testing with hypothesis

Most tests currently use handpicked input values and assert specific outputs. Hypothesis generates inputs automatically and finds edge cases you wouldn't think to write. E.g. https://github.com/ecmwf/anemoi already uses it for testing attention and MLP layers.

Candidates

• Standardization: output should always have mean ~0 and std ~1 regardless of input shape or value range
• Clamping: model output should always stay within declared bounds for any input tensor. test_clamping.py already tests specific cases, hypothesis would test the invariant more thoroughly
• Dataset slicing: test_time_slicing.py hardcodes specific input/output pairs - these could be expressed as properties that hold for any valid combination of ar_steps, num_past_forcing_steps, num_future_forcing_steps
• Datastore interface contracts: get_dataarray output shape should always match num_grid_points for any valid split and category

Setup

hypothesis would be a dev dependency only and integrates directly with pytest via @given decorators, so no CI changes are needed.