fix(bridge): gate batched-list position_ids on the target model

[sohv]

Fixes #1626.

Batched list input builds an attention_mask and position_ids itself so pad tokens don't contaminate the forward. The mask is safe for any model, but the position_ids were handed over unchecked — a forward taking neither position_ids nor **kwargs raises TypeError where it would otherwise have returned logits.

This is the gap that was raised while reviewing #1610. That PR added _accepts_derived_position_ids() and gated the main forward() derivation, but we left these two sites for a follow-up because I couldn't produce a model that demonstrably failed there. The LLaDA test harness builds a fixed-signature forward in-process, which reproduces it:

TypeError: TinyLLaDAModelLM.forward() got an unexpected keyword argument 'position_ids'
kwargs attempted: ['attention_mask', 'position_ids']

What I changed

Both sites now call the same _accepts_derived_position_ids() helper, so there's no new predicate. The attention_mask stays unconditional -it's safe for every model, and withholding it would reintroduce the padding contamination the branch exists to prevent.

Verification

One regression test in the LLaDA suite, red on dev-4.x with the TypeError above and green with the fix. It asserts both halves: no position_ids reaches the model, and the attention_mask still does. A single unbatched string was never affected and is covered as a control.

Full unit + integration + acceptance: 6594 passed, 1 failed. The failure is test_bridge_hooked_parity_multi_step_optimization, which fails identically on unmodified code and is --ignored on the macOS CI job. mypy clean.