[BUGFIX] Use the rendered width for srcset descriptors

[wazum]

Variants were labelled with the requested width, so a candidate the processor could not deliver at that size still claimed it. With processor_allowUpscaling off, an 800px original asked for five variants produced five candidates pointing at the same file under five different descriptors, and the browser dropped all but one of them as repeated width descriptors.

The descriptor now comes from the processed image, and a width that is already in the srcset is skipped. Where the requested width can be delivered, which is the normal case, the output is unchanged.

The functional test runs with upscaling disabled and uses its own fixture rather than storage_with_file.csv, because it needs a sys_file_metadata row: a processed file that falls back to the original carries no dimensions of its own and reads them from the original's metadata.

Fixes: #103

[achimfritz]

👍 , thanks