Cosmetic: Fix the checked-out branch badge being dim and never updating - #697

Make the sidebar's checked-out badge show the state the window is actually
in, rather than the one it happened to be in when the row was built, and
give its two states a look that tells them apart.

How We Got Here

Found while chasing why the UI screenshot comparison reports differences
on pull requests that change nothing visible (#693). Two back-to-back
runs of the screenshot tests, same commit and same repository state,
produced a badge that was grey in one and blue in the other.

• Refresh the badge when its window becomes or stops being the main
  window. Nothing in the app observed window state at all, and the badge
  is a bitmap baked at the moment the row is configured. The first
  sidebar reload runs from awakeFromNib, before the window is main, so
  the badge baked grey and stayed grey until some unrelated reload
  happened to rebuild it. Which side of the shutter that landed on is
  the difference above.
• Give the checkmark the highlight color when its row is emphasized in
  an active window. A row is only emphasized while its window is key, so
  the key-window branch this replaces could never be taken by the
  selected row and the checkmark was always dimmed.

The two belong together: after the second change isMainWindow is the
only window signal the control reads, so the observers added by the
first cover all of it.

Beyond the screenshots, this is visible in ordinary use: the badge does
not dim when GitX goes to the background, nor light up when it comes
back, for any row that is not the selected one.

Making The Inactive State Actually Look Inactive

Fixing the refresh exposed a second problem: the badge now updated on
losing focus, but barely changed when it did. The active and inactive
capsule colors turned out to have the same brightness, 0.3972 against
0.3987 in relative luminance, so going inactive changed only the hue.

• Darken the inactive capsule, a 22% drop in relative luminance, so an
  inactive badge reads as dim rather than merely as less blue.
• Knock the checkmark out of the capsule while the window is inactive,
  leaving it hollow so the row behind shows through it. Painting the
  checkmark the sidebar's own color was the obvious alternative and is
  wrong: the sidebar is 246 and a selected row is 220, so a fixed color
  matches on an unselected row and reads as a pale checkmark on a
  selected one.

The two are linked. The hole is only visible while the capsule differs
from what shows through it, which puts a floor under how light the
capsule can go. Matching the capsule to the selection color, which would
be the natural next step, makes the badge vanish outright on a selected
row: same color, and a checkmark that is a hole rather than a shape.

Expected Behavior (and Look)

1. Click a branch row in the sidebar.
   The checked-out row's capsule turns white -
   the checkmark inside should be blue, not grey.
   
2. Click into the history list.
   Badge returns to blue capsule with white checkmark.
   
3. Switch to another app, then back.
   The badge should dim and re-light. (Master does neither.)
   The capsule goes to a darker grey and the checkmark becomes a hole.
   
4. Select a different branch so the checked-out row is unselected, then
   switch apps again (the case AppKit never rescued, where Master is
   most obviously stuck).
   

The images in 3 and 4 were taken before the hollow checkmark and the
softer capsule, so they show the right states in the older colors.

Test plan

• Every change here was run with the fix stashed out and again with it
  applied. In each case only the case under test failed without it,
  while the controls passed in both runs.
• make unit-test passes: 158 tests, no failures.
• The color tests use a window double reporting itself main, a test
  process having no way to make a window genuinely main; the refresh
  tests assert the badge is rebuilt rather than what color it becomes,
  for the same reason.
• The dim and hollow states are asserted on measurements rather than on
  fixed colors: one test computes relative luminance and requires the
  inactive capsule to be at least 20% darker, another renders both
  badges and requires the inactive one to carry more transparent pixels.