Remove n300-llmbox from perf benchmarks

[vkovacevicTT]

Ticket

Closes #4949

Problem description

We have successfully enabled benchmarks on qb2-blackhole.
n300-llmbox runners produce very long queue times and should be deprecated.

What's changed

• Removed n300-llmbox from benchmarking CI.
• Transfered vllm n300-llmbox benchmarks to qb2.
• Added qwen_3_32b_tp_qb2 to OnPR uplift CI, both perf and accuracy benchmark.
• All prior llmbox models now run on qb2.
• Enabled opt-level=2 on vllm_falcon3_7b and vllm_falcon3_10b that were previously failing on llmbox:
  Enable opt-level 2 for vllm_falcon3_7b_tp (fails with L1/circular-buffer clash) #5438
  Enable opt-level 2 for vllm_falcon3_10b_tp (fails with L1/circular-buffer clash) #5439

Checklist

• qb2-blackhole

  • llama_3_1_70b_tp PCC failure - issue llama_3_1_70b_tp PCC error on QB2 #5487
  • vllm_gemma4_31b_it_tp L1 OOM - issue Enable opt-level 2 for vllm_gemma4_31b_it_tp (fails with L1 out-of-memory) #5440

• Sanity check

[vvukomanTT]

Great, this will help reduce queue times massively!
One note, let's remove all references for n300-llmbox tests then

tt-xla/.github/workflows/pr-main.yml

Line 83 in 9b0c875

{ "runs-on": "n300-llmbox", "filter": ["test_llama_3_1_70b_tp"] } ]'

also has some, maybe in other places too.

[nsumrakTT]

I love this PR!
Thank you!

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 33.85%. Comparing base (b621dab) to head (72df121).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5436   +/-   ##
=======================================
  Coverage   33.85%   33.85%           
=======================================
  Files          37       37           
  Lines        4989     4989           
=======================================
  Hits         1689     1689           
  Misses       3300     3300           

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