[FEATURE REQUEST] Optimize ImageInput::read_tiles for multi-part EXR

In the case of EXR tiled multi-part files, ImageInput::read_tiles will almost never fall in the case
```
if (native_data && contiguous
&& (xend - xbegin) == nxtiles * spec.tile_width
&& (yend - ybegin) == nytiles * spec.tile_height
&& (zend - zbegin) == nztiles * spec.tile_depth)
```
because it only works if the image has size multiple of tile size.

Then the code goes into ImageInput::read_tile which forgets completly about the requested number of channels and ends up decoding all the channels, just to return at the end the channels that were requested.
For high resolution EXR files with a lot of channels, this makes openimageio compltely unviable when encoutering tiled EXRs.

Could this case be adapted to just decode the requested channels in the first place?

Trace ImageInput::read_tiles and ImageInput::read_tile, focusing on the multi-part tiled EXR path and requested-channel handling. Compare the fast-path conditions with the fallback behavior; done means the fallback avoids decoding unrequested channels while preserving correct tile reads.

Written by the indexing model from the issue text.