Follow-up to a76d14e: update enhanced drawing flag
from 2 to 0x11, as 0, 1, 2 and 3 might be used for
categorising crengine internal bookmarks.
This might avoid breaking upstream/other client code,
which could then just add 'flags &= 0x10' somewhere to
benefit from enhanced drawing.

So it gets a fixed id and we get more cache stability.

By storing an order sequence number in the _specificity field,
which will ensure selectors with the same specificity are checked
and applied in the order we parsed them.

Embedded floats overflowing their containing paragraph, and
not directly children of that final node (so, wrapped by more
inlines nodes), would not be visited by elementFromPoint()
and text or image inside them would not be found.

Made the 'direction' argument to elementFromPoint() and
createXPointer() more explicite, and allow finding the first
or last word on a line in logical-order.
This allows not missing text and links in pages where the
first or last line is RTL or bidi.

Previously, images were incorrectly skipped in the avoidWrap
checks. So, a sequence of word-space-image-period was handled
as word-space-period, and a wrap avoid before a period forced
the whole thing to be on a same line.

Will allow a style tweak to confine text to its paragraph
line height (aka the strut):
- this will prevent sub- and superscript-like inline elements from
  increasing line height and pushing down the baseline, and will
  reposition them if needed to limit their overflow.
- this will resize inline images so they are not taller than the
  line height.

Makes thing easier to read and understand.
Also use proper lInt16 for it, as it can be negative.

Positive text-indent should apply only on the first line of
a final block (paragraph), and should not apply after a <BR>.
(Same for negative indent, that should apply on all but first
line, including the first line after a <BR>.)
Properly handle negative indents (that induce a left shift of
the first line in LTR, possible overflowing its container),
and hanging indents (that induce a right shift of the non-first
lines in LTR, without any overflow), both in RTL a LTR cases.
(Note that the text-indent "hanging" it not handled by all
current browsers, but we do, as it is used in fb2.css.)

From upstream in-progress "merge with KOReader" PR125.

Some word segment might start with a non-text source, without
a ->t.font field, which could cause a segfault when accessed.
Very rare and hard to reproduce, so did not try to make that
skipping more part of the whole flow (we should probably also
not provide them to HyphMan::hyphenate() to not miss some
hyphenation opportunities).

Also renamed inner 'start' and 'end' vars to avoid shadowing
upper ones with the same names, to avoid confusion.

Since f3b5958 and possibly looking for hyphenation on more
than one word word, we may set hyph flags at multiple positions,
and have non last words larger than they should, causing the
appearance of a space. So, only add the hyphen width on
the last word of a line.
Noticed when hyphenating "extre<SPAN id=z/>mement", hyphenable
as "extre-me-ment" and chosing the 2nd one: it would be rendered
as "extre me-" (with a spurious space) on first line and "ment"
on next line.

If a paragraph starts with a quotation mark or dash, and is
followed by a space (classic or nbsp): don't have this space
width widened or shortened in the process of text justification.
This generalizes the existing code that was doing this only
for 4 kind of dashes: we do it for more opening marks, dashes,
bullets, and other ascii chars that could start a line.