BUG: Static decoding is from left to right with temp=0

[JacobHelwig]

There is a bug in the code that makes static decoding with temp=0 or top-k=1 always decode the left-most token. This bug also affects dynamic decoding.

First we process logits. With temp=0 or k=1, this turns them into 1-hots. This makes the probability of sampled tokens at all masked positions 1.

Raw logits per masked position (illustrative):
  pos 0:   [..., -2.1,  5.3,  1.2,  4.8,  3.1, ...]
  pos 1:   [..., -1.0,  0.2, -3.0,  5.5,  2.1, ...]
  pos 2:   [...,  3.1,  1.1,  4.0,  2.8,  4.5, ...]

After top_k=1 (non-argmax logits → -inf), then softmax:
  pos 0:   [...,  0.0,  1.0,  0.0,  0.0,  0.0, ...]   ← one-hot
  pos 1:   [...,  0.0,  0.0,  0.0,  1.0,  0.0, ...]   ← one-hot
  pos 2:   [...,  0.0,  0.0,  0.0,  0.0,  1.0, ...]   ← one-hot

  → seq_x0_p = gather(probs, sampled_token) = 1.0 at every masked position.

When static sampling, this means we call top-k on a tensor of all ones. To break the tie, we select the first (left-most) position.

Decoding a 4-token block over 4 steps (num_to_transfer=1 each step):

Step 0: block state = [MASK,  MASK,  MASK,  MASK ]
        confidence  = [ 1.0,   1.0,   1.0,   1.0 ]    ← all tied
        topk(k=1) on tied 1.0s → picks index 0 (PyTorch tiebreak)
        commit at pos 0
                ↓
Step 1: block state = [tok_0, MASK,  MASK,  MASK ]
        confidence  = [-inf,   1.0,   1.0,   1.0 ]
        topk(k=1) → pos 1
                ↓
Step 2: block state = [tok_0, tok_1, MASK,  MASK ]
        confidence  = [-inf,  -inf,   1.0,   1.0 ]
        topk(k=1) → pos 2
                ↓
Step 3: block state = [tok_0, tok_1, tok_2, MASK ]
        commit at pos 3.

  → every block is decoded strictly left → right, regardless of which
    positions the model is actually most confident about.

When dynamic decoding with temp=0, since the threshold condition is applied to the processed probabilities (all 1 still), it will result in all positions satisfying the condition if threshold < 1, and thus, all positions are decoded.

Decoding a 4-token block, dynamic_threshold = 0.9:

Step 0: block state = [MASK, MASK, MASK, MASK]
        confidence  = [1.0,  1.0,  1.0,  1.0 ]
        confidence > 0.9 = [True, True, True, True]
        → COMMIT ALL 4 POSITIONS IN ONE STEP

  → every block collapses to a single forward pass; each position is
    sampled in isolation, block-diffusion's iterative refinement is
    skipped, and the output is typically garbled, repetitive text.

These bugs are in the original SDAR code, but they have since started using a corrected sampling pipeline.