Add MiniMax-H3 Turbo video generation support

[TroyHernandez]

MiniMax-H3 Turbo looks like a good diffuseR target for the video path.

Model/repo:

• https://huggingface.co/lightx2v/Minimax-h3-Turbo
• https://github.com/ModelTC/Minimax-H3-Turbo
• Base model: https://huggingface.co/MiniMaxAI/MiniMax-H3

Why it fits:

• The HF repo is tagged for Diffusers and Image-to-Video.
• It covers text-to-video, image-to-video, and reference-to-video style workflows.
• The Turbo checkpoint distills MiniMax-H3 into 4-step and 8-step inference, which is the practical path for local use.
• The base H3 model is large enough that diffuseR should treat this as a video/runtime capability target, not as an image-model variant.

Suggested diffuseR surface:

download_minimax_h3_turbo()

minimax_h3_turbo(
  prompt,
  image = NULL,
  last_image = NULL,
  reference = NULL,
  steps = 4,
  resolution = "544p",
  output = "output.mp4"
)

Implementation notes:

• Keep this in video-specific files, e.g. R/minimax-h3.R / R/recommend-video.R, so it doesn't tangle the image-model core.
• Add recommend("minimax-h3-turbo") tiers once we know real VRAM/RAM behavior.
• Prefer explicit download consent and artifact checks, consistent with existing model download policy.
• Start with the Turbo repo rather than base H3. Base H3 is a 33B dense video/audio model and likely needs a very different deployment story.
• Check whether the Diffusers path already exposes all modes we care about, or whether the first diffuseR pass should use a small Python bridge/container-style runner.

Acceptance sketch:

• download_minimax_h3_turbo() fetches/verifies the needed artifacts.
• A smoke example can generate a short low-resolution clip from a prompt or image.
• recommend() returns a non-CPU plan on hardware that can realistically run the Turbo checkpoint, and clearly refuses hardware that can't.