fix: align eval user_id default and support recursive evalset discovery

[brucearctor]

Summary

Fixes two server-side issues from #5423:

1. Align default user_id fallback (test_user_id → user)

When an eval set doesn't specify a user_id, the eval runner was defaulting to "test_user_id", while the adk-web frontend defaults to "user". This mismatch caused session 404 errors when the UI tried to fetch eval result sessions.

Changed in:

• evaluation_generator.py
• local_eval_service.py

Note: The companion frontend fix (using the userId from the eval result response instead of hardcoding) is submitted separately to google/adk-web#424.

2. Warn on undiscoverable evalset files in subdirectories

LocalEvalSetsManager.list_eval_sets() only scans the immediate agent directory. .evalset.json files in subdirectories (e.g., app/eval_sets/my_test.evalset.json) are silently ignored, causing confusion.

Now logs a clear warning message when subdirectory evalset files are detected, telling the user to move them to the root directory:

WARNING - Eval set file 'eval_sets/nested.evalset.json' found in subdirectory and will be ignored. Move it to '/path/to/app/' to make it discoverable.

Note on port change (Issue 3): adk web now serves on port 8000 instead of the previous 8501. This is not a bug but should be called out in migration/upgrade notes.

Testing

• All 30 existing + new tests pass in test_local_eval_sets_manager.py
• Added test_local_eval_sets_manager_list_eval_sets_subdirectories to verify the warning log behavior
• Updated list_eval_sets_success test to use real filesystem instead of mocking os.listdir

Related

• Issue: Evals UI: session 404, non-clickable results link, undocumented port change, unclear evalset discovery #5423
• Frontend companion PR: google/adk-web#424