fix(core): a graded clip's canvas no longer comes out blank on a real GPU - #6615

What changed

A graded video could show as a flat background instead of its footage in the preview: the grade drew a blank canvas and hid the source video under it. Two causes, both in the runtime's colour grading:

1. A clip shown far past the screen asked for a canvas the GPU cannot draw. The canvas was sized from the clip's on-screen box, so a clip zoomed to about three times the frame asked for 6582 x 11520 pixels. The GPU silently clamped the shared WebGL drawing buffer (to 4353 x 7620 here), and every later copy out of it, which reads from the bottom of the buffer, landed outside what had been drawn. The canvas is now capped at 2^24 pixels and 8192 on a side, in the clip's own shape, and a draw whose drawing buffer is smaller than its canvas gives the clip back to its source video (shown again, canvas hidden) and throws instead of copying, including for a clip that had already drawn.
2. The first draw of a video could sample nothing. loadeddata and seeked arrive before the GPU can upload the frame, so that draw produced a blank canvas, and nothing redrew until the next seek. A video now also redraws once its frame is presented (requestVideoFrameCallback, a single draw when paused).

What I measured

Real GPU: Chromium (Electron) on an Apple-silicon Mac, a 9 s montage with several graded clips, the preview page served by hyperframes preview with the runtime swapped for each build. After the page loads, the playhead goes to 0.4 s, 0, 0.4 s, 0, and each shown clip's grading canvas is read back.

Three runs after the change gave the same readings. With the change, the frame at 0 matches hyperframes snapshot, which shows ungraded footage there.

What I did NOT exercise

• CI has no GPU. The tests below run against the mocked WebGL context, so they pin the size cap, the refusal to hide the source over a clamped buffer, and the draw on a presented frame, not the clamp itself.
• Software GL (SwiftShader) never showed the problem, so a green run there says nothing about it.
• Only a Mac GPU; other GPUs clamp at other sizes, which the 2^24 cap sits under but I did not measure.

fails without the fix: caps a clip shown far larger than the screen to what a GPU can draw, keeping its shape, leaves the source showing when the drawing buffer cannot hold the canvas, gives a drawn clip back to its source when a later draw outgrows the drawing buffer, and draws a paused video again once its frame is presented, after the first draw sampled nothing in packages/core/src/runtime/colorGrading.test.ts; all four fail on main and pass here (113 of 113 in the file, three runs in a row).