Bugs to fix:
(not including ones for async, for those see below)

?* Handling of general entities in non-expanding mode wrong:
  eats entities (finishCharacters() buggy?)

Fixes to conformance:

[DONE] * Coalescing mode
[DONE] * Do NOT report ignorable white space in prolog, by default
    (won't implement XMLInputFactory2.P_REPORT_PROLOG_WHITESPACE for now)
[DONE] * EventWriter should force non-lazy parsing mode for stream reader.
    (ditto for SAX reader)
* Support for entities (explicitly defined ones, XInclude,
  or DTD-induced)
* Basic DTD handling? Starting with the internal subset:
  * Entity value parsing? (no PE expansion, yet)
  * Attribute type info?
  * Attribute default values?
[DONE] * Repairing writer (non-repairing implemented)

Testing:

* Create basic event (start/end elem/doc, characters, CDATA, PI, comments)
  tests over all scanners; including async one (need faced to abstract
  out differences). This because StaxTest generally will only test one
  or at most two types of backend XmlScanner impls.

Async:

* Multi-byte char handling not implemented for comments
* Handling of:
   * PI
   * CData
   * Entities (partial impl, not with blocking yet)
* Bootstrapping (xml decl handling)
* API for constructing instances (AsyncInputFactory?)
* API for feeding more data

Performance:

[partial] * Inline 2-byte UTF-8 handling for text? (to help w/ russian, german)
  [note: doesn't seem to help!]
[DONE] * Separate char type codes for tables, to make valid codes consequtive
  and minimize value space (may be faster)
* XmlScanner.bindNs(): replace linear scan with hash, for big tables
 (test with degenerate doc like nux-samples/ns-soap.xml!)
* Namespace URIs:
    * Can we avoid intern()ing/canonicalizing URIs? (not without changing 
      comparisons). If we can, only intern() first N distinct (to avoid
      degenerate cases; like 16); check canonical for M (like 64)
[DONE] * Specialize InternCache used for URIs, not to require construction
      of Strings; use a non-linear hash (since URIs are loooong)
* Prefix String creation/interning: optimize? (since prefixes tend
  to cluster)
  (note: similarly for output side?)
[Done?] * For output side: figure out how to cache serialized names
  (just the last one?)
[DONE] * Optimize parsePName() of StreamScanner, for the common case where
  buffer does have 8 more bytes: inline handling of first 4 or 8 bytes.
* Utf8XmlWriter.doConstructName(): replace calls to String.getBytes()
  with more efficient (for short names) version