DCTDecode (!!)

main high level optimizations to be done:
	- AttributeName: don't return a string but a unique TSymbol (how to make it so? The name would have to be qualified a lot) or intern the string.
fix memory leaks:
	- fix reference cycles ("/Parent" etc)
cleaning up:
	- get rid of TSymbol.
	- get rid of the interfaces.
	- get rid of all the Variant typed attributes that you can.
	- have one hash table per PDF that lists all the known objects (and indexes them by key) (and owns them).
	- PDFs.TPDF.Symbols : modify type to a normal hash table with objects as value (no interfaces, no nothing).
extra features:
	new-style PDF1.5 XRef "object" (Mechanics).
main program:
	check for ItalicAngle and output <i></i>
decoder:
	fall back to descendant font for FontDescriptor.