Remove CUDA from benchmarks directory.

[ysiraichi]

This PR removes the CUDA specific code from the benchmarks directory. This is in line with the CUDA deprecation that started on release 2.8.

Key Changes:

• Removed the keep_model_data_on_cuda parameter
  • Used in combination with zero-overhead CUDA to XLA:CUDA data movement, removed in #9598 and #9603
• Deleted llama.py, nightly.sh, run_benchmark.sh, run_single_graph_bm.sh, and run_top_tier_bm.sh
  • All of them ran benchmarks comparing PyTorch Inductor with XLA:CUDA, specifically

[zhanyong-wan]

Impressive!