Resolves #1000.  In Flot 0.7 we only calculated tickDecimals and
tickSize once, when creating the tickGenerator for the first time.  This
meant that calls to setupGrid failed to recalculate the values, as
reported in #860.  #861 fixed the problem by moving calculation into
tickGenerator, but this turned out to cause a new problem, since the
function doesn't run when an explicit ticks array is provided.

This commit solves both by always recalculating the values outside of
the tickGenerator function.  As far as I can tell the only reason it
wasn't done this way from the beginning was to avoid unnecessary work in
the case where tickGenerator is already provided in the options.  But
the extra work is negligible, and it's actually more consistent for the
properties to always be set.

Ole's original implementation used 'middle', which I switched away from.
After a great deal of testing it turns out that 'middle' does in fact
provide the most consistent results, so we're switching back to it.

The values don't change, so there's no reason to repeat those
calculations on every redraw.  The resulting code is not just faster,
but also smaller and simpler, and we no longer need to store halign in
the text info object.

The setting is the same for everything, so there's no reason to re-set
it every time.

The core implementation used parseInt, which was originally to catch
text values, but really only stayed due to its side-effect of flooring
the values.  The canvas implementation has never rounded coordinates.
This led to various one-pixel rendering glitches between the two
implementations.  I've fixed the problem by consistently rounding
coordinates to the nearest whole number.

Line height is a core font property - it can be specified as part of the
unified 'font' definition - and we therefore need to give users access
to it when they're manually defining the font.

This way it doesn't inherit from something further up, whose line-height
may be more or less than we expect.

The way in which xaxes/yaxes inherit options from xaxis/yaxis resulted
in a minor bug, where tickColor defaulted to the xaxis/yaxis color
instead of the color for its axis.  Fixed by applying the default before
extending the per-axis options, resolving #984.

There's still some questionable behavior here; this section should be
revisited for 0.9, especially with an eye towards removing some of the
code that only exists for backwards-compatibility.

This resolves #991, replacing the earlier temporary patch.  It takes
advantage of the fact that line-height can take the form of a unit-less
integer, in which case it mirrors the font-size, even when it is
something abstract, like 'smaller'.  We can then read the dummy
element's height to learn the effective font-size.

The purpose of the stylesheet hack was to provide a default without
having to use inline styles on containers.  We can do this much more
neatly by instead just giving the inline styles to a parent container,
leaving users free to customize the children.

This partially addresses #991, but a more complete solution, or better
documentation, is still necessary.

There's no need to continue in the function when all highlights are
being cleared.

The unhighlight method retrieved the point differently from highlight,
and incorrectly.  It now matches the behavior of highlight.

Add an extra pixel to ensure that the lines cleanly join both top and
bottom ticks.  Also apply the sub-pixel offset only in the necessary
direction, so the lines don't start a pixel off from where they should.

The axis color now controls the color of the axis line, instead of its
ticks and labels, while the tickColor controls the tick color. This
makes a little more sense and provides the minor feature of being able
to change the axis line color separately from that of its ticks.  Pull
request #917 ought to be easier to merge now, too.

Plugins are re-initialized with each re-plot (which may not be the right
thing to do, but that's how it works for now).  The previous approach of
saving references to the original Canvas functions therefore broke,
since the second time around we'd get a reference to our new function.
Instead we hold those references as globals within the plugin, and only
set them once.

This whole idea of replacing prototype functions is, now that I step
back and look at it, really awful.  This needs to be changed ASAP to
something less ridiculous.

This allows us to reuse the canvas elements without having to retain the
rest of the Canvas object, which should really be reset when the plot is
reconstructed.  It's also a little simpler, and the Canvas constructor
just feels like the right place for this code.