Fix text prefiller start pos being updated twice

[pytorchbot]

In text_prefiller.cpp prefill() we are updating the start_pos inside prefillChunk() as well as outside. This PR fixes this issue. Also updating llama main.cpp to take max_new_tokens.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15531

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• ROCm failures during provisioning step due to network issues

❌ 1 New Failure, 2 Cancelled Jobs

As of commit 9f74ef6 with merge base e0dda90 ():

NEW FAILURE - The following job has failed:

• pull / test-eval_llama-wikitext-linux / linux-job (gh)
  RuntimeError: Command docker exec -t a0555b6cb8448a1e0be872c346441a2cc7419de699e12d905365f080a94b9e2f /exec failed with exit code 1

CANCELLED JOBS - The following jobs were cancelled. Please retry:

• pull / unittest / macos / macos-job (gh)
  ##[error]The operation was canceled.
• pull / unittest-editable / macos / macos-job (gh)
  ##[error]The operation was canceled.

This comment was automatically generated by Dr. CI and updates every 15 minutes.