Replace assert statements with proper exceptions for runtime validation

[RajdeepKushwaha5]

Description

Several places in the codebase use assert statements to validate user-provided inputs (config values, dataset coordinates). Python's assert statements are disabled when running with the -O (optimize) flag, which means these checks silently disappear in optimized deployments — potentially leading to data corruption or cryptic downstream errors instead of clear, early failures.

Affected locations

File	Line	Statement	Risk if skipped
mllam_data_prep/ops/selection.py	L70	assert sel_start != sel_end	Equal start/end silently creates an empty slice
mllam_data_prep/ops/selection.py	L82–84	assert len(ds[coord]) > 0	Empty coordinate selection goes undetected
mllam_data_prep/ops/cropping.py	L56	assert da_lat.dims == da_lon.dims	Mismatched lat/lon dims produce a wrong convex hull mask
mllam_data_prep/ops/cropping.py	L57	assert da_lat_ref.dims == da_lon_ref.dims	Same for the reference dataset
mllam_data_prep/config.py	L494	assert args.f.endswith(".yaml")	Non-YAML file silently passes validation

Proposed fix

Replace each assert with an if … raise ValueError(…) that includes a descriptive message with the actual values involved. For example:

# before
assert sel_start != sel_end, "Start and end cannot be the same"

# after
if sel_start == sel_end:
    raise ValueError(
        f"Start and end of range selection for coordinate "
        f"'{coord}' cannot be the same (got {sel_start!r})"
    )

The same pattern already exists elsewhere in the codebase (e.g. the raise ValueError checks in selection.py L63–65 and stacking.py L31–34), so this would make error handling consistent.

Motivation

• python -O is commonly used in production and containerised deployments for performance.
• The Python docs explicitly warn against using assert for data validation.
• The sister repo neural-lam tracks an equivalent issue (neural-lam#278).

Scope

This is a small, focused change — 5 assert → if/raise replacements, plus corresponding tests to verify that a ValueError is raised in each case. Happy to submit a PR if this sounds reasonable.