Support scalar results in combined evaluations

[zanvari]

Summary

Fixes #516.

Some evaluation modules (e.g. wer and cer) return scalar values rather than dictionaries. CombinedEvaluations._merge_results() assumes every module returns a dictionary and attempts to call .keys() on each result, causing combined evaluation to fail.

This change wraps scalar results using the corresponding evaluation module name before merging, allowing scalar-returning evaluation modules to be combined while preserving the existing behavior for dictionary-returning modules.

Tests

Added a regression test covering:

evaluate.combine(["wer", "cer"]).compute(...)

which now returns the expected combined evaluation results without raising an exception.