Fix eval max length

[Some-random]

Although in theory we should allocate a big buffer for reasoning models to play with before giving out final answers, in practice, our generation can never be that long (in fact, the max_completion_length parameter for model training is usually clipped at 1024). So it doesn't make sense to use 32768 as the max length for generation.

Changing 32768 to 4096 would also resolve some issues because 32768 is too long as the default length for some models