Jev-like mode for DiffusionGemma model (github.com)

The recent update to the DiffusionGemma model introduces a "Jev-like" mode aimed at enhancing the model's functionality for AI developers. This mode implements three new optional fields in the sampling parameters: `diffusion_seed_canvas`, which allows users to define specific template slots along with positions for model completion; `diffusion_max_steps`, specifying the maximum denoising steps before final commitments; and `diffusion_read_only`, which enables immediate logging of outputs upon reaching the maximum steps. These changes aim to provide more structured and controlled responses from the model, addressing prior issues of mismatched log probabilities and incomplete outputs. This update is significant for the AI/ML community as it enhances the usability and performance of the DiffusionGemma model by making it easier to manage and set conditions for requests. The introduction of checking mechanisms for seed lengths and token caps helps prevent errors that may lead to performance degradation or crashes. Overall, these improvements not only streamline the process for developers but also bolster the reliability of generative tasks performed by the model, paving the way for more complex and refined AI applications.