[gemm3] Fix EOS token handling for Gemma text-only export

[seyeong-han]

Summary

When exporting Gemma models with --task "text-generation", the C++ text runner wasn't stopping at the <end_of_turn> token (ID 106), only at the <eos> token (ID 1). This caused generation to continue beyond expected stopping points.

This diff modifies save_config_to_constant_methods() to export multiple EOS token IDs via get_eos_ids, which the C++ runner's kEosIds method already supports.

Changes

• EOS list handling: Build a list of EOS token IDs that handles both single int and list cases from config.eos_token_id
• Gemma detection: Detect Gemma models via config.model_type containing "gemma" and automatically add token 106 (<end_of_turn>) to the EOS list
• Dual export: Export both get_eos_ids (full list) for C++ runner compatibility and get_eos_id (first element) for backward compatibility

Compatibility

• Backward compatible: Python modeling.py already checks for both get_eos_id and get_eos_ids
• C++ runner ready: The C++ runner's get_eos_ids() function already supports reading a list of EOS tokens via the kEosIds method

Test Plan

• Export Gemma-3-1b-it with --task "text-generation" and verify metadata contains get_eos_ids: [1, 106]
• Run the C++ text runner and verify generation stops at <end_of_turn>
• Run existing tests: pytest tests/models/test_modeling_gemma3.py -v

[seyeong-han]

I believe this is not a universal way to adopt the various chat-templates.
So, I created another PR which can accept the chat-template variations by utilizing jinja format.

pytorch/executorch#16987