Tune quickstart capture framing and downsampling for documentation media

[mataylor-nvidia]

Description

The quickstart media generator rendered the OVRTX path tracer straight to the 320x240 publication size. Robot silhouettes and shadow edges aliased badly at that size, and the per-task focal lengths left the subject occupying a small fraction of the frame with a large empty floor area around it.

This changes the capture and encode stages only:

• Supersample. Capture at 1280x960 and downsample with lanczos in the generator, instead of path-tracing directly to the final size. This is what recovers the drawer handles, robot edges, and shadow definition.
• Framing. Tighten the focal length per task to the largest value that still keeps the subject inside the published crop.
• Denoise. Ease hqdn3d from 2:2:6:6 to 1:1:3:3. The heavy setting existed to hide path-tracer noise; supersampling now removes that noise without blurring away detail.
• Palette. Widen from 128 to 256 colors and switch bayer to sierra2_4a dithering, which removes banding on the smoother gradients the sharper frames produce.

Before / after

End-to-end effect on a published frame — current pipeline on the left, this PR in the middle, and a 640px variant on the right that was evaluated and not taken:

The framing options that were compared before settling on the tighter value:

Full variant sweep, including rtx_environment="studio", which was rejected — it blows out the background and lowers contrast rather than raising it:

Per-task focal lengths

The safe focal length is scene-specific rather than a fixed ratio of the old value. Scaling every task by the same factor clips two of them, so they are tuned individually:

Task	Before	After	Why
Franka cabinet	22	32	Most headroom of the four; subject roughly doubles in frame
G1 locomotion	20	29	Verified fully in frame
Cartpole	18	21	26 clipped the pole out of the crop; 23 grazed the top edge
Kuka Allegro	22	24	32 clipped the arm, and 26 left little headroom. Held back because the lift policy raises the arm above the pose a zero action holds

Cartpole and G1 under the naive proportional scaling — note the Cartpole pole leaving the top of the crop:

Cartpole sweep that established 21 as the safe maximum:

Kuka Allegro sweep showing the arm clipping at 32:

Each value was validated by capturing the task and inspecting the frame through the same crop and downscale the generator applies.

Review screenshots are hosted on the media-review-assets branch of the author's fork so they are not committed to this PR.

Media refresh

This PR changes tooling only. Since #7501 moved the quickstart GIFs to S3, docs/source/setup/quickstart.rst points at download.isaacsim.omniverse.nvidia.com URLs while the generator still writes agent-comparison.gif and task-sampler.gif into docs/source/_static/quickstart/. Regenerated media has to be uploaded to the bucket separately before readers see any change.

Type of change

• Tooling / documentation media quality

Checklist

• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation — n/a, no documented API or page content changes
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works — n/a, the generator has no test harness; validated by rendering each task and inspecting the published crop
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file — n/a, no source package changed
• I have added my name to the CONTRIBUTORS.md or my name already exists there