XML code completion improvements

XML code completion improvements

This changeset contains a number of improvements to the code
completion in XML files.

(1) Add suffix completion in attribute values. For dimensional
    attributes, this will add in (or complete prefixes of) units like
    "dp", "sp", etc. For fractional attributes, the % and %p suffixes
    are completed. (In both cases, there is also a tooltip message
    provided shown in the completion documentation popup).  In
    addition, for flag values, when completing at the end of a flag,
    then the separator character (|) along with the other flag values
    (except those already set) are offered.

(2) Handle completion when the caret is not at the end. For example,
    if you have the following source:

        layout_w^idth="wrap_content"    (^ = caret position)

    then if you invoke code completion you will also be offered the
    attribute layout_weight, and selecting it will -replace- the
    layout_width attribute with layout_height, it will not insert
    layout_weight in the middle of layout_width. This handling is done
    not just for attribute names as shown here but for attribute
    values and element tag names as well.

(3) Improve @resource/ completion sorting. When completing resources,
    consider which attribute is being completed, and use that
    information to sort the resources.  For example, if completing a
    "text" attribute then @string/, and @android:string/ resources
    will be listed first.  For attributes like margins @dimen will be
    first, for include layout the @layout attributes will be listed
    first, and so on.  Also, the resources are sorted alphabetically
    as the second sort key.

(4) Hide some builtin Eclipse code completion items. In particular,
    the default code completion provider for Eclipse will always add
    the current attribute value as a completion. This is extremely
    confusing since if you have a typo, code completion will tell you
    that what you have is okay. I personally witnessed this confusing
    a user who had typed something from memory and code completion
    seemed to him to "confirm" that he had it right.

    To fix this, the code which installs completion providers, will
    identify the builting WST completion provider, and when found,
    replace it with a "filtering" wrapper. This completion provider
    delegates all its calls to the WST completion provider, but it
    recognizes a few patterns, in particular the above attribute value
    completion proposal as well as some namespace and schema ones, and
    removes these.

    (The reason we only filter out these items instead of removing it
    completely is that the completion provider also appears to be
    responsible for inserting code templates defined by the user, and
    we don't want to neuter those if defined by the user.)

(5) Fix a bunch of corner cases - flags could only have a single
    separator (|), resource attribute completion items were missing
    icons, and completion in some positions was not working.

(6) Unit tests. There are now comprehensive unit tests for code
    completion. Each unit test points to a particular source file
    and a particular position within the source file, and code completion
    is invoked for that position. A set of unit tests dump out the
    allowed completion items for each such position, and these are
    compared against golden files. A second set of unit tests then
    specify a particular code completion item among the choices to
    be applied, and then applies that completion item to the document.
    This, along with the caret position (indicated as ^) is written
    into a golden file and used for comparisons. To make it easy
    to see what these tests do, the golden file is actually recorded
    as a diff, so in the test you will see something like this:

Code completion in completion1.xml for android:gravity="left|b^ottom"
selecting bottom:
<         android:gravity="left|bottom"
---
>         android:gravity="left|bottom^"

(6) There are some other misc changes. We had some code which used
    "dip" for device indepdendent pixels (such as the AbsoluteLayout
    handler). Since "dp" seems to be preferred, the code will now
    use that instead.

Change-Id: I8bd5c8336d8747dac1f10a9269ea4197f304cb70