When the web script is reactivated, refreshKeyGrabs is called, but
locusOfFocus might not yet be set to the currently active control, so the
script doesn't know that structural navigation is active, meaning that at-spi
isn't instructed to add the key bindings for structural navigation. Later,
onFocusedChanged is called, and it might update locusOfFocus without notifying.
We need to ensure that the AT-SPI key grabs are updated in this case.

Fixes issue #289.

According to the ARIA spec, the tab role's children are presentational
and therefore user agents should not expose its children. Chromium
follows this; Firefox does not. As a result, we risk double-presenting
the label of a page tab in Firefox when in browse mode. Fix this by
enforcing the children-presentational within Orca.

Fixes issue #288.

This should get the original, intended functionality working again.

See issue #47.

KP_Delete is by default bound to an Orca command. If that command is
unbound, we should treat KP_Delete just like Delete in terms of speaking
the new location.

If a single space (e.g. in a paragraph) separates two other objects
(e.g. links), instead of speaking the space we might speak a different
part of that same text element. The reason why is that currentLineText
prefers the existing substring if it exists (in this case it does, the
space). But in the case of non-editable content, it then strips that
substring and only returns the result if a result remains (it no longer
does). If getting the substring fails, we fall back on getTextLineAtCaret
which can lead to repetition of text.

Stripping the substring was a carry-over from old Gecko support and has
been around for 8+ years, likely to work around some oddity in Gecko's
accessible text implementation. Gecko's accessible text implementation
has improved quite a bit since then, so with any luck stripping the
string is no longer needed. Therefore, solve the repetition by no
longer stripping the string.

A side effect of the above fix was that we would speak "space" for the
space. This is due to speechdispatcherfactory's _speak calling
adjustForPronunciation which, among other things, converts a single
character (i.e. to be spoken alone) into its character name. In the
case of SayAll, we're not really speaking the space alone and we don't
want to make that adjustment. So have adjustForPronunciation not turn
single characters into their names during SayAll.

Fixes issue #273.

The ancestry of newly-selected page tabs typically is flagged as layout
only and is thus not presented by Orca. If, however, the ancestor has
a name, it is no longer considered layout only. Firefox's page tabs are
contained in a named toolbar. As a result, when the active browser tab
is changed via keyboard shortcut, we present not only the name of the
now-active/selected tab but also the parent toolbar.

There's probably no good reason to ever present the ancestry of a
newly-selected page tab, so enforce that.

See issue #276.

columnHeaderForCell can now return None in the case of error conditions.
Thus we need to check for that in the generator method.

See issue #274

See issue #271

Fixes recent regression presenting newly-edited calc cell

There are several scenarios we need to handle when a page loads and
focus changes:

1. The document is loaded in the same browser tab, resulting in the
   current caret context being null/dead. This causes us to get the
   new caret context for the document.
2. The document is loaded in a new tab, and the user agent has updated
   the accessible embeds relation. This causes us to not yet have a
   stored caret context for the document, so we get it.
3. The document is loaded in a new tab, but the user agent hasn't yet
   updated the accessible embeds relation and the old document is also
   opened. Here we think the user is still in the old document, and it
   has a valid caret context. Thus when the user arrows in the newly
   loaded page, we present the page from the previous tab.

Address the third issue by checking if the document changed. If it
has, clear the existing caret context and search the event source
which emitted the focus event for the new context.

Fixes issue #264.

Fixes issue #270.

When we use presentObject, the formatting string usually includes
the label and/or name. But it does not for all roles, such as the
PUSH_BUTTON role. Using presentMessage with the event.any_data (if
present) is consistent with what we do for description changes and
will ensure the name change of the currently focused/active item
is spoken.

See issue #255.

This change retrieves all the headers for a cell for implementations
which do the following:

1. Only give us the innermost/closest header for a cell
2. Support returning the header of a header

Firefox is such an implementation; Chromium is not. While Chromium does
give us all the headers for a cell (good; nothing to work around), it
does not support returning the header of a header. Thus when navigating
in a nested row/column header, we cannot present the changed outer
header. That should be fixed in Chromium.

Fixes issue #269 (to the extent we can do so in Orca)

Because row/column headers might themselves have headers.

Orca now prefers the table cell interface over the table interface to
get a cell's position. Gecko implements the table cell interface, but
instead of returning (row, col), it returns (col, row). This can cause
Orca to (among other things) not announce row and/or column headers.
See https://bugzilla.mozilla.org/show_bug.cgi?id=1794100

Work around that by not using the table cell interface for cell position
in Gecko apps until the bug is fixed and has been fixed for a while.
(E.g. in an ESR release)