[Bug] Recovering logic of a long evicted request is broken

[masahi]

https://github.com/octoml/mlc-llm/blob/batch-serving/serve/mlc_serve/engine/engine_common.py#L385-L399

For streaming case, we cannot clamp the generated tokens and recompute them.
Moreover, since the clamping logic is done in the worker but not in the main process, the discrepancy arises between the main and the worker process. See #158 and #164.

We need to either

• Require that generation never grows beyond max_num_batched_tokens
• Split recovering of such requires into multiple batch, using the new evaluate_multi_query function from Add new Relax function to the batched model for evaluating query tokens over multiple time steps in parallel  #156

@elvin-n @sunggg

[masahi]

@elvin-n After #157 lands, you can follow a similar strategy to use multiple EvalMultiQueryRequest to split restoring of a long request into several batches, each of which fits into max_num_batched_tokens.