RenderTarget: Fix copy of images. - #30585

…ot landing tonight

Kevin authorised landing the upgrade if it works ("if the upgrade works, then do
it"), then went to sleep. It does not clear the bar I set before measuring, so it
is reverted and this is the record instead.

MEASURED, with a presence control, production bundle, identical 12s probe:
  three 0.172.0 + postprocessing 6.39.1  ->  6 glBlitFramebuffer messages
  three 0.174.0 + postprocessing 6.39.4  ->  0

The control is the point. A zero from an instrument never shown to produce a
non-zero is not evidence, and this repo has shipped that mistake seventeen times.

That confirms the upstream maintainer's theory on pmndrs/postprocessing#750: the
GL storm we had filed as a SEPARATE production-bundle finding and the GodRays
regression share a root cause in three r172's depth-texture handling
(mrdoob/three.js#30585, fixed r174).

A gap this exposed, worth more than the upgrade: prod-smoke.mjs CANNOT SEE this
error class. It filters on m.type() === 'error' and these arrive as 'warning', so
the repo's own production gate is structurally blind to the error storm it exists
to watch.

WHY IT IS NOT LANDING: r174 changes the whole look. 19 of 31 frames go over gate,
many at 99%+ local density, and the sky is visibly more saturated -- a global
colour/tone change, not one subsystem regressing. Landing it means rewriting all
31 baselines, which is a judgement about how the game should LOOK. That is
Kevin's call, not a correctness fix the loop can self-gate, and a bundled
engine-bump + re-baseline is precisely the shape baseline-trailer exists to stop.

It also runs against c1d933e, which deliberately blocks 0.x MINOR bumps for the
engine stack. This measurement is evidence FOR that policy.

Reverted fully: package.json + package-lock.json restored from backup, npm ci
re-run, 0.172.0 / 6.39.1 confirmed installed, tree clean.

The upstream reply is DRAFTED, NOT POSTED -- a public upstream comment is
outward-facing and stays Kevin's. It says plainly that I cannot yet confirm r174
fixes the SUN, because the global change makes that unseparable without a
controlled r174 pair I have not run.