Batch generation can silently omit prompts and mishandles max_tokens=0

[Thump604]

Problem

Two generation boundary cases fail unclearly:

1. A per-prompt list such as max_tokens that is shorter than the prompt list is consumed with unchecked zip, silently omitting trailing prompts.
2. stream_generate(max_tokens=0) skips the generation loop and later reads locals that were never initialized, raising UnboundLocalError.

Expected behavior

Per-prompt option lengths should be validated before generation, and a zero-token request should terminate normally without entering token generation.

Proposed fix

Draft PR #1681 validates per-prompt list lengths and handles the zero-token path explicitly, with focused regressions for both cases.