[LLM runner] Correct the deprecated seq_len CLI help text

Problem

The llama runner help for --seq_len says to use max_seq_len, but there is no corresponding CLI flag. The implementation falls back from --max_new_tokens to --seq_len, and its runtime log tells users to prefer --max_new_tokens.

This mismatch was found while validating LLM documentation in #22620. The C++ source change is intentionally excluded from that documentation-only PR.

Proposed change

Update the --seq_len help text to recommend --max_new_tokens and clearly distinguish total generation length from new-token count and the model context limit.

Test plan

Build the llama runner and inspect --help; run existing llama runner CI.

cc @larryliu0820 @cccclai @helunwencser @jackzhxng @digantdesai