Add DiffusionGemma (text): block-diffusion model + diffusion_generate

[gilbert-barajas]

DiffusionGemma (text) for mlx-lm

A text-only port of DiffusionGemma (google/diffusiongemma-26B-A4B-it) — Google's
encoder-decoder block-diffusion MoE built on the Gemma 4 stack. Addresses the request
in #1391 (text serving in mlx-lm; mlx-vlm already covers the multimodal path).

Instead of emitting tokens autoregressively, the decoder denoises a fixed 256-token
canvas over reverse-diffusion steps (entropy-bound acceptance + renoise, temperature
schedule, adaptive stop), with the prompt prefilled into an encoder KV cache.

Credit

This mirrors @Blaizzy's working mlx-vlm DiffusionGemma4Backbone 1:1 (so the existing
mlx-community/diffusiongemma-* conversions load directly, no reconvert) and uses the HF
transformers reference for the sampler semantics. Grateful for both as parity targets.

What works

• Full encoder-decoder model + a top-level diffusion_generate (a sibling to the AR loop, since
  the canvas-denoising contract doesn't fit generate_step).
• Loads the existing mlx-community/diffusiongemma-26B-A4B-it conversions (4-bit and bf16)
  with zero key/shape mismatches — sanitize drops the vision tower + handles the MoE
  weight names; from_dict flattens the nested text_config.
• CLI: mlx_lm.generate dispatches to diffusion_generate when the model exposes it
  (generic hasattr check; no model-specific import in generate.py).

Verification

• MoE numerically parity-checked vs transformers: Router top-k indices exact, Experts
  max abs diff ~1e-9 (machine precision).
• Adversarial code review against the HF + mlx-vlm reference caught and fixed two sampler
  bugs (committing the argmax canvas vs the renoised one; carrying temperature-scaled vs raw
  logits for self-conditioning).

Receipts (Apple Silicon, M5 Max)

Running the real 26B-A4B through this port (256-token canvas, cold):

build	size	tok/s
4-bit	16.5 GB	~250
bf16 (full precision)	51.6 GB	~121

mlx_lm.generate --model mlx-community/diffusiongemma-26B-A4B-it-4bit \
    --prompt "Why is the sky blue? Answer in one sentence."
# -> "The sky is blue because Earth's atmosphere scatters shorter wavelengths
#     of light (blue and violet) more effectively than other colors."

Coherent on prose and code (an is_prime with the 6k±1 rule).

Status / remaining (hence draft)

• Single-canvas generation (≤ canvas_length); the >256-token block-autoregressive outer
  loop is a follow-up.
• mlx-lm-format unit tests + a logits-parity test to add.
• Open shape questions for maintainers (from my Model type diffusion_gemma not supported. #1391 comment, still unanswered): is a separate
  diffusion_generate the seam you'd prefer, or dispatch inside generate? Appetite for
  mlx_lm.server support here vs a follow-up? Any preferred encoder-decoder cache convention?

Happy to iterate on the shape — opening this as a draft to move the conversation from the
issue to actual code. Closes #1391 when complete.