Preserve exact-zero SFT self-conditioning

[sylvesterkaczmarek]

Addresses the self-conditioning semantics reported in #773.

Summary

Preserve an exact-zero self-conditioning embedding in the Hackable Diffusion SFT adapter when self-conditioning is disabled.

• represent training-time self-conditioning enablement with an explicit per-example mask instead of replacing logits with zero logits
• make the first SFT decoder pass produce a true zero self-conditioning embedding
• mark diffusion sampling step 0 as non-self-conditioned and later steps as self-conditioned
• add regression coverage for exact-zero and mixed enabled/disabled self-conditioning states

Why

The current adapter represents “no self-conditioning” using all-zero vocabulary logits and then passes them through encode_logits. Because encode_logits applies a softmax and embedding projection, zero logits produce a vocabulary-mean embedding rather than an exact zero self-conditioning signal.

The main DiffusionGemma sampler already initializes self-conditioning with an all-zero embedding tensor. This change makes the Hackable Diffusion SFT path consistent with that behavior without changing the meaning of actual previous-pass logits.

This intentionally addresses only the self-conditioning portion of #773; it does not change the separate multi-canvas cache/decode behavior discussed there.

Testing

• targeted Hackable Diffusion SFT/state-handler tests: 38 passed
• pyink --check passed on all five changed files