feat: move circular padding to per-generation params

[fszontagh]

Summary

Moves the circular_x / circular_y flags from sd_ctx_params_t to sd_img_gen_params_t and sd_vid_gen_params_t. Previously users had to unload and reload the model to switch between tileable and non-tileable output; now they can vary it per generate call on a persistent context.

The setter set_circular_axes is just two bool assignments on the runner context, and the VAE encode/decode entry points already take circular_x / circular_y as function arguments, so nothing about the plumbing required a load-time decision.

Related Issue / Discussion

None.

Additional Information

Behavior verified

Loaded AcornXL_V2 (SDXL) once, ran two consecutive generations of "seamless tiling brick wall texture, flat lighting" at 512x512, 20 steps, karras, seed 42:

• With --circular: log shows Using circular padding for convolutions (x=true, y=true), output has wrapped edges.
• Without: log stays silent, output is a normal brick wall.

Same executable, same code path, no reload.

Changes

• sd_ctx_params_t: dropped circular_x, circular_y (default init, param string, and print). Load-time set_circular_axes calls in StableDiffusionGGML::load_from_file removed.
• sd_img_gen_params_t, sd_vid_gen_params_t: appended circular_x, circular_y; wired defaults in _init and printing in sd_img_gen_params_to_str.
• New helper apply_circular_axes_to_diffusion(sd_ctx, cx, cy) invoked at the top of generate_image / generate_video. Sets the sd struct fields (so the existing VAE guard picks them up) and calls set_circular_axes on diffusion_model, high_noise_diffusion_model, and control_net. Logs only when a flag is on.
• CLI: --circular, --circularx, --circulary moved from SDParams to SDGenerationParams; assignment routed to sd_img_gen_params_t / sd_vid_gen_params_t in to_sd_img_gen_params_t and to_sd_vid_gen_params_t.
• Enum values and existing gen struct field order preserved; new fields appended.

Notes

• No CUDA / Vulkan gating - circular padding is model-side conv padding, backend-agnostic.
• The VAE tiling adjust logic (configure_image_vae_axes) continues to read sd_ctx->sd->circular_x/y. Those are now set from gen params at generate-time, one line before the guard runs.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.