Feature/jan/decoding performance

[Schnjan]

• Since encoded blocks usually contain a lot of empty values, we now utilize this information to skip large parts of the decoding calculation
• The discrete cosine transform now stores the results in a way so we only need one loop to iterate
• The block metadata is now loaded from the data array using spans instead of a binary reader
• There's a new shortcut to extract single slices

[daniel-flemming]

You might also create a data structure holding these 3 buffers - the code would be easier readable and you can move the Rent/Return logic into a method of that structure.