QoL: Default to using fast tokenizer for Llama models

[arnavgarg1]

Quality of life improvement - we used to set use_fast to False since the tokenizer was not natively written so it would take 2-4 minutes to load, however in newer transformers versions there is first class support so we no longer have to do this.

[github-actions]

Unit Test Results

  4 files  ±0    4 suites  ±0   31m 57s ⏱️ + 1m 59s
31 tests ±0  26 ✔️ ±0    5 💤 ±0  0 ❌ ±0 
62 runs  ±0  52 ✔️ ±0  10 💤 ±0  0 ❌ ±0 

Results for commit 14b3534. ± Comparison against base commit 42723e3.

[justinxzhao]

Thanks for following up on this! I was wondering about this and why setting use_fast=True didn't seem faster.