Problem: `read_exr_channels` kept the exr crate's whole decoded image alive
while it built every f32 plane, and copied each channel through an extra
`Vec` (`values_as_f32().collect()`) before placing it. One decode of a
2560x1440 Blender file with 32 channels briefly held ~320 MB of samples next
to ~470 MB of planes, per concurrently decoded frame.
Fix: consume the decoded image channel by channel, so each channel's samples
are freed (or moved) as its plane is made. A channel that covers exactly the
display window is converted in one pass, and an f32 one (Cryptomatte ranks)
becomes its plane without a copy; other channels are placed into the display
window with bounds-checked indexing, as before. Keeping half channels as f16
and sizing the channel cache by bytes stay open.
Test: a full-window f32 channel's plane is the same allocation; f16 / u32
channels and offset or clipped data windows are placed as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
7aee1d6275