Honor the model's generation_config.eos_token_id in the transformers backend

[waqaskhan137]

Fixes #1278.

What

For generative tasks, _generate_padded's generation_config.update(...) sets eos_token_id=self.tokenizer.eos_token_id, overriding the terminators the model itself declares in generation_config.eos_token_id. Chat models whose turn terminator is not the tokenizer's eos never stop: they emit their turn-end token, it is ignored, and every generation runs to max_new_tokens.

Concrete case in the issue: Gemma ends chat turns with token 106 while tokenizer.eos_token is <eos> (id 1); the model declares generation_config.eos_token_id = [1, 106, 50]. With the override, an MMLU chain-of-thought task padded every generation to the cap, with 63-95% of returned tokens being token 106 repeated.

Fix

Prefer self.model.generation_config.eos_token_id when set, falling back to self.tokenizer.eos_token_id (one line, plus a comment). Models whose declared eos already equals the tokenizer's are unaffected.

Validation

Measured on google/gemma-4-E2B-it, MMLU CoT item, RTX PRO 6000, bf16 (details in #1278):

	before	after
tokens_generated	7168 (= cap)	2654 (stops on its own)
token-106 padding	6774	1
latency / item	145 s	55 s
extracted answer	correct	correct (unchanged)

ruff check and ruff format --check pass on the touched file.