Eval: loss-only path + compute receipt

[StanByriukov02]

I was evaluating GENIE-style checkpoints and kept running into the same thing: sometimes I only need CE loss (teacher-forced logits), but the current eval loop still does MaskGIT refinement + decode/LPIPS/acc work.

This PR adds a clean fast path for that case.

What changed

• --loss_only: compute CE loss from temporally teacher-forced logits and skip MaskGIT sampling (no refinement loop), skip decoding/LPIPS, and skip accuracy.
• --skip_decode: keep sampling/logits, but skip decoding + LPIPS (useful if you still want generation metrics without the heavy decode pipeline).
• --device {cpu|cuda}: makes it easier to reproduce runs on CPU-only machines too.
• Receipt: prints compute_logits_calls and compute_logits_calls_per_frame, so you can see how many full forward-logits passes actually happened.
• xformers becomes optional: if it’s not installed (or XFORMERS_DISABLED=true), the code falls back to the basic attention path.

Guarantees / scaling / what I’m not claiming

• Guaranteed GPU-compute reduction for --loss_only: per generated frame, full forward-logits passes drop from ~maskgit_steps to 1. For maskgit_steps=2 that’s ~2× on the forward-pass-heavy part.
• Scales with quality mode: if someone runs maskgit_steps=K, the compute ceiling is ~K× (so 5–10× is possible at K=5..10).
• CPU side matters too: skipping decode/LPIPS/acc avoids a big chunk of CPU + pipeline overhead when you only care about loss, and the new receipt makes it obvious you’re not doing extra work.
• Not promised: end-to-end Joules/frame for the full eval script (depends on dataset IO, decode, etc). I did run a small forward-pass microbench on an H100 to sanity check direction.

H100 sanity check (forward-pass microbench)

• maskgit_steps=2, baseline does 2 compute_logits calls per timestep, loss_only does 1
• measured ratio_J ≈ 1.83× lower Joules for the forward-pass segment (same model config, repeated loop to smooth power sampling)

Repro (receipt)

• baseline:
  python genie/evaluate.py --checkpoint_dir --maskgit_steps 2
• loss-only:
  python genie/evaluate.py --checkpoint_dir --maskgit_steps 2 --loss_only

You should see compute_logits_calls_per_frame drop ~2.00 -> ~1.00 for maskgit_steps=2.