More line breaking fixes

See merge request !507

Some of the layout tests now get better line breaks.

It turns out that we do need to look for longer breakpoints
after all if we want to find optimal solutions.

When looking for breakpoints, we were assuming that
there is no point to keep looking forward once we
hit a spot where the broken off part is too long.

But with hyphen insertion, that is no longer true,
necessarily. Consider 'bli '. Breaking after 'bl'
will insert 'bl-', which might be longer than
'bli', which is what will be inserted when we
break after the i.

To fix this, keep looking for breakpoints as long
as there is still a chance to find one.

We always need to check if we still fit, after shaping.
The width we use before is just an estimate.

The key insight here is that if we find a broken item
does not fit, we are only interested in finding an
*earlier* breakpoint - the later ones aren't going
to produce a shorter run.

So we can just keep track of the last char we want
to allow breaking at.

This fixes the case of finding the minimum width
with wrap mode PANGO_WRAP_WORD.

Fix a bug in Small Caps handling

Closes #627

See merge request !505

We were not passing the correct split_offset to
pango_item_split().

Testcase included.

Fixes: #627

Simplify breakpoint disabling

See merge request !504

This keeps spiraling :(

If the break is at the end of the item, we
were forgetting to check for the space before
the break.

When we are breaking after a space, we must not
count the width of the space towards the line,
since we are zeroing it later.

It is a bit annoying that there are multiple places
where this has to be taken into account.

Another missing bit in this code is that we are
only looking at a single whitespace character before
the break, when we should really look for a sequence
of spaces.

When unsplitting an item, we were sometimes
calculating the remaining_width incorrectly.

We can use the log_widths that we already have.

We only want this inside process_item, so we can
make this a purely local thing, without modifying
log_attrs.

Cosmetics

See merge request !503

coretext: correctly clamp the core text weights at min/max values

See merge request !502

Addresses bug found in #624

layout: Try harder to not produce overlong lines

See merge request !501

Our accounting for run lengths is imperfect (mainly
due to log widths for clusters being evenly distributed),
so it can happen that after reshaping the split item,
we find that it does not actually fit in the remaining
width. Previously, we would just use the split run
at that point and produce an overlong line. Instead,
undo the split, disable the breakpoint we used, and
try again.

This undoes the effect of pango_item_split.
We will use this in future commits.
Private for now.

This will let us use some of them during layout.

GTK frequently uses this width during measuring,
so it is good if we can reproduce what happens
in this case.

This can make it easier to reproduce problems.

layout: Fix a poblem with hyphen width

See merge request !500

This checks behavior that is required by GTK.

In some cases, we were forgetting to account
for the width of a hyphen.

Mention new attribute values in markup docs.

Bump fontconfig dep to 2.13.0

See merge request !499

Call hb_font_set_ptem when creating fonts

See merge request !498

We are requiring a new enough fontconfig now.

This version of fontconfig has all the API
we are using, and was released in 2018.