Perform explicit culling step only if LLG is submitted

[fjetter]

This explicit culling step is for many reasons redundant.

1. DataFrame expressions behave well and generate only what's needed. Legacy Array collections perform this as a low level optimization, etc.
2. The Scheduler._generate_taskstate method is already walking the graph in a way that automatically culls. Therefore, this optimization step is only there to cull the graph before it enters dask.order

I believe the only path where this even has a chance to trigger is if the Client.get receives a raw dictionary and in these cases we have zero information about what happened to the graph. It still makes sense to cull to avoid worst case runtimes but for all other cases this isn't necessary, I believe.

[github-actions]

Unit Test Results

See test report for an extended history of previous test failures. This is useful for diagnosing flaky tests.

    26 files   -     1      26 suites   - 1   10h 56m 24s ⏱️ - 29m 16s
 4 106 tests ±    0   3 988 ✅  -     2    112 💤 ±  0  6 ❌ +3 
48 809 runs   - 2 670  46 614 ✅  - 2 551  2 189 💤  - 121  6 ❌ +3 

For more details on these failures, see this check.

Results for commit 4378728. ± Comparison against base commit 6691e27.

♻️ This comment has been updated with latest results.