Bound stop-token check to written tokens in dflash_generate

[SuperMarioYL]

Problem

dflash_generate pre-allocates output_ids with mask_token_id past the
prompt at dflash/model.py:79. The in-loop early-exit check at the bottom of the
decode loop scanned the full pre-allocated tail:

if stop_token_ids is not None and any(
    stop_token_id in output_ids[:, num_input_tokens:]
    for stop_token_id in stop_token_ids
):
    break

When mask_token_id happens to be one of the stop_token_ids — a
model-config-dependent edge case the project already cares about
(see #76 "Preserve output tokens that equal mask_token_id") —
mask_token_id in the unwritten tail of the buffer satisfies the
in check on the very first iteration and generation aborts after
one block.

Fix

Aligning with the post-loop trim a few lines below — which already
uses torch.isin over the trimmed slice — the in-loop check now
scopes the scan to output_ids[0, num_input_tokens : start + 1],
i.e. positions that have actually been written this run. The
pre-allocated stop_token_tensor is hoisted out of the loop so both
the in-loop and post-loop checks share it.

Tests

Added tests/test_model.py covering pure-Python / pure-tensor logic
that runs on CPU without weights:

• build_target_layer_ids interpolation (1-/2-/4-layer cases)
• extract_context_feature offset+concat shape and values
• sample argmax / temperature paths
• regression test for the buffer-scan pattern: reproduces the
  legacy check firing spuriously, asserts the new check does not
• sibling test confirming the new check still fires on a real
  stop token after the cursor advances

Wired in via a [project.optional-dependencies] test extra so
existing backends are unaffected:

uv pip install -e ".[test]"
python -m pytest tests/test_model.py -v
# 8 passed in 9.15s

Refs #76.