feat(rollout): forward --sglang-text-encoder-precisions to the engine

[Rockdu]

Based on refactor/fsdp-5-precision-plans (radixark#106); retarget to main once that lands.

What

Forward --sglang-text-encoder-precisions from miles to the rollout engine.

Why

text_encoder_precisions is a PipelineConfig field, so miles' auto-prefix already exposes --sglang-text-encoder-precisions on the CLI, but _compute_server_args only forwards dit_precision / vae_precision explicitly. Passing the flag today is silently dropped and the engine keeps its per-pipeline defaults. This bites whenever the DiT precision is moved off the pipeline default: SD3 ships ("fp16", "fp16", "fp32") text encoders, so an fp32 DiT run dies in the engine with mat1 and mat2 must have the same dtype until the encoders follow.

set_defaults(..., None) keeps "unset" detectable, so a run that does not pass the flag forwards nothing and the engine's own defaults still win.

Files

• miles/utils/arguments.py — default the passthrough arg to None
• miles/backends/sglang_diffusion_utils/sglang_diffusion_engine.py — add the field to the explicit forwarding tuple

Checklist

• pre-commit run --all-files — not run locally
• Added/updated tests — none; this extends an existing forwarding tuple with no new branch
• pytest -x — not run locally
• python3 train_diffusion.py --help parses, and an SD3 GRPO run launches with the flag set (verified on an H200 devbox)
• CLI reference docs — n/a, no new flag
• Example walkthrough — n/a

[Rockdu]

Moved to radixark — a PR inside the fork is not reviewable there.