Clean-up CUDA cache.

[ysiraichi]

This PR clears CUDA cache on 2 cases:

• benchmark instance was deleted (i.e. it was first instantiated with cuda as device, but it's not needed anymore)
• inner model has to be moved to an XLA device
  • since Re-land: Fix model initialization. #6296, models are first initialized in the original accelerator, and then moved to XLA device

Additionally, it also fixes a bug where the instantiated benchmark wouldn't be deleted if it didn't have a default precision set (in default_precision_flag property).

[vanbasten23]

Sorry for the late review. What issue do we get if we don't clean up the CUDA cache? @ysiraichi

[ysiraichi]

I was getting OOM when training llama_v2_7b_16h with both inductor and openxla backends.