Stop the frozen-generator check from rejecting inner-graph inputs

[jessegrabowski]

The check that refuses a graph whose generator nothing advances counted any non-output use as a draw. An op carrying an inner graph might draw from a generator it's handed or might just accept it, so a layer that takes one without drawing got rejected with a message claiming the graph draws from it. Now only direct draw ops count — inner graphs are collect_default_updates' business, and it already raises when one draws without threading the generator back out.

Also restores the "here's how to fix it" half of the two inner-graph error messages, dropped when we vendored them from pymc.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.73%. Comparing base (5860ec9) to head (e78adbd).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #87      +/-   ##
==========================================
+ Coverage   95.46%   96.73%   +1.27%     
==========================================
  Files          45       45              
  Lines        1807     1807              
==========================================
+ Hits         1725     1748      +23     
+ Misses         82       59      -23     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.