fix: update report merging logic to handle non-unique method names

[tlambert03]

Report.from_directory collects reports in a dict keyed by the name stored inside each JSON file. run_evaluation.py stores every method's report under the same name (default_report), so with the documented layout

reports/DDeepLift/default_report.json
reports/DIntegratedGradients/default_report.json

each file overwrites the previous one and merge_reports is handed a single report. The result is one method's report, not the per-sample best across methods (with no error). This means the per-image choice between Discriminative
IG and Discriminative DeepLift described in the paper never happens.

Checklist

• I have added tests to show that my feature works in various situations.
• I have tested the updated pipeline on an example dataset.
• I have included docstrings
• I have updated the How-To in the documentation to reflect the changes
• I have updated the Changelog to reflect any changes to the default function of the pipeline