Fix PPL VLM last-position scoring: rewind-free incremental teacher forcing

[amd-genmingz]

Summary

• Rewrites the last-position-only branch of compute_sample_nll in
  tests/PPL/perplexity_vlm.py to score targets without rewind_to.
• Uses a single generator with incremental teacher forcing: prefill the
  prompt once, then append_tokens the true previous target token per
  step and read the final-position logits. generate_next_token is never
  called, so no sampled token is inserted into the sequence.

Why

The previous last-position path relied on rewind_to + append_tokens to
recover per-position logits. On gemma4 multimodal + amdgpu EP with
past_present_share_buffer=false, the OGA per-layer KV-cache branch does
not maintain the sequence-length bookkeeping (shape_[2]) across a rewind.
After a rewind, OGA's pre-allocated present.* shape disagrees with the
EP-computed shape (e.g. {1,8,277,256} vs {1,8,296,256}), and the native
layer fatally aborts with 0xC0000409 (STATUS_STACK_BUFFER_OVERRUN).

What changed

• Drop all rewind_to calls from the last-position branch.
• Prompt prefill (and the vision encoder / image features) now runs once;
  each target position costs a single incremental decode step. Cost goes from
  O(target_len) full prefills back to O(1) prefill + O(target_len) cheap
  single-token decodes.
• The all-position (if) branch is unchanged; models that emit full
  [B, T, V] logits still take the single-pass fast path.
• Dump fields (target_ids / target_nll / top1_ids / topk) are
  unchanged, so downstream compare_vlm_ep_cpu.py is unaffected.
• Minor: fix argparse help string (%% escaping and en-dash normalization).

Correctness

• After set_inputs(prompt), final-position logits = P(? | prompt) =
  prediction for target[0].
• After append_tokens([target[j-1]]), final-position logits =
  P(? | prompt + target[:j]) = prediction for target[j].
• Mathematically equivalent to the original rewind logic and to the
  all-position path (same set of target-position NLLs).