DIS: run the variational refinement on the whole pyramid

[qwertypower]

The variational refinement ran once, at the finest level, after the coarse-to-fine loop had already finished. OpenCV's DIS runs it inside that loop, at every level, and the difference is not cosmetic: every level was handing the next finer one a raw densified patch flow as its initialisation instead of a refined, coherent one, so patch-search errors compounded all the way down the pyramid and the single pass at the end had to undo the lot at the most expensive resolution.

Each refinement intermediate becomes a mip chain with a view per level, so the VR shaders keep deriving their extent from imageSize() and are unchanged. The dispatch block moves into the level loop as dis_vr_level(), and the patch search now initialises from view_flow_refined[l + 1] rather than view_dense[l + 1].

On a level that is not getting the full solve, prep and add still run: prep zeroes dW as a side effect, so add stores flow_refined = dense exactly. Two dispatches and seven taps a texel keep the descriptor wiring static - flow_refined is always the valid flow for its level - which is much cheaper than switching descriptor sets between refined and raw sources per frame.

How many levels get the full solve is a tier on the generation count, beside the iteration counts already there. The flow chain runs once per SOURCE frame, so at x4 the guest sits at a quarter of the panel rate and the chain runs a quarter as often: half the load per second of the x2 case, which is the heaviest the system ever gets. x2 therefore keeps the refinement to the finest level and is bit-identical to before; x3 and x4 spend the headroom that is already there on the whole pyramid, and that is exactly where the flow is worst because the motion between two source frames is twice as large.

Modelled per source frame at the Balance preset (448x252), against today's x2 load as the ceiling:

x2   30 fps guest   33.0M taps   +0.6%   0.99G/s   101% of ceiling
x3   20 fps guest   45.3M taps  +11.8%   0.91G/s    92% of ceiling
x4   15 fps guest   48.2M taps  +12.9%   0.72G/s    74% of ceiling

Descriptor pool: the VR sets now exist once per level, 10 per level instead of 10 total. Pool and demand are both derived from the same constants and match exactly at 227 sets / 1375 samplers / 307 storage images (was 157 / 815 / 167). An undersized pool here fails with OUT_OF_POOL_MEMORY and frame generation silently stops existing, which is why it is computed rather than written out.

Memory cost is the extra mips on nine scratch images: +33% of their level-0 footprint, about 3.4 MB at the Balance preset.

Also points the DEVAR credit in the README at https://devar.ai/.

Claude-Session: https://claude.ai/code/session_01MCkAQJH8ik5iJZqf2NX3w6