Quantization fixes

[milesial]

• enables FP8 KV quant (k and v scales, no q or prob)
• fixes the default saved max model len (256k)
• use static FP8 activation scales instead of dynamic (tensor-wise, minmax) for better perf
• direct copy of tokenizer and preprocessing configs to avoid llmcompressor introducing silent truncation

[spectralflight]

Thanks! A few minor comments, but otherwise LGTM. Please run just lint per https://github.com/nvidia-cosmos/cosmos-reason2/blob/main/CONTRIBUTING.md#test