Z-Image-Turbo text-to-image support

[TroyHernandez]

Adds Z-Image-Turbo (Tongyi-MAI, Apache-2.0) text-to-image: a 6B
guidance-distilled single-stream DiT that renders legible text in
images (English and Chinese), 8 steps, no CFG.

download_zimage_turbo()   # ungated; ~33 GB download, one-time fp8
                          # quantize to a 5.9 GB artifact
txt2img_zimage("A sign that reads \"DIFFUSER\" carved in wood",
               seed = 42)   # or txt2img(..., model_name = "zimage")

Architecture (ported from the diffusers reference, Apache-2.0)

• Transformer: single-stream joint attention (no MMDiT split);
  30 trunk blocks + 2 modulated noise-refiner + 2 unmodulated
  context-refiner blocks; sandwich RMSNorms (learned norm before AND
  after attention/FF); scale/gate-only modulation with tanh-squashed
  gates (no shift); SwiGLU with separate gate weights; 3-axis RoPE
  theta 256; sub-sequences padded to multiples of 32 with learned pad
  tokens. flux_attention reused as-is (bias = FALSE).
• Text encoder: the FLUX.2-klein Qwen3-4B port reused verbatim
  (tokenizer files are byte-identical between the checkpoints); only
  the chat template differs (enable_thinking = TRUE, no think block)
  and the pipeline takes the penultimate hidden state, mask-sliced to
  the caption length.
• VAE: the FLUX.1 16-channel decoder unchanged (the checkpoint's
  VAE config is the flux-dev autoencoder).
• Scheduler: FlowMatch static shift 3.0 on sigmas
  linspace(1, 1/N, N); the model consumes the reversed normalized
  timestep (1000 - t)/1000 and its output is negated before the Euler
  step.
• Quantization: one-time local fp8 (or NF4) through the existing
  family-hook machinery; cast set is the 7 core linears x 34 blocks =
  238 weights, adaLN/embedders stay bf16. The fp32-shipped checkpoint's
  pass-through residents are stored bf16.

Allocator gc-gate fix (cross-model)

Generation initially ran at ~142 s regardless of resolution; profiling
showed 89% of wall time in R gc (~2,200 gcs x 62 ms). torch's
allocated/reserved ratio gate (0.95) is chronically exceeded under
backend:native (steady state 0.957), firing the R-gc callback on
nearly every CUDA allocation — and start_torch() reads the gate
options only once at init, so mid-session option changes are inert.
.flux_gc_gates() now pushes thresholds into the live allocator:
ratio gate off, absolute allocated_rate 0.65 as the
garbage-accumulation gate, footprint-based reserved_rate as the net.

Measured 1024x1024 on the RTX 5060 Ti 16 GB:

Model	Before	After
Z-Image-Turbo	143 s	24 s
FLUX.2 klein	48 s	13 s
FLUX.1-schnell	121 s	55 s

Verification

• Per-module parity fixtures against the diffusers reference: position
  ids bit-exact, full tiny-model forward < 1e-4, strict bidirectional
  key censuses, chat-template token ids exact against the shipped
  tokenizer (CJK/emoji included), quantized round trips (NF4 + streamed
  and resident fp8). Full suite: 692/692.
• AST-based structural audit of the translation (numeric literals +
  call names per paired scope): clean; only dead defaults and
  omni-only constants differ.
• Full-size load census exact (238 cast weights); same-seed output
  reproduced identically before/after the gate fix.
• R CMD check clean.

Sample outputs in ~/Sync (zimage_sign_1024.png shows the text
rendering).