After experiments with loading the whole file at once, and
with decoding and parsing in separate thread, lazy reader
turned to be the fastest/least memory intensive solution.
Characters are now decoded in small batches.
This improved parsing speed by ~20%.
No global state anymore. Anchors are now zero terminated strings
and TagDirectives are a simple array. Event structure was changed
to prevent size increase.
Minor fixes and improvements.
Removed endMark from Node to keep it in 32 bytes on 64bit.
This will result in slightly worse debugging messages, but we
still have the start position of a node.
Tag is needed for better compliance with the spec and emitting
support for multiple tags with the same D data type.
