refactor(usage): extract the usage-percent widgets onto a shared module

[zachthedev]

Now that #456 has merged this is a single standalone commit against main.
The extraction sits on top of that PR's formatPercent(value, format) threading
rather than colliding with it, which is why it was sequenced this way.

Five widgets carried five copies of one render body: SessionUsage.ts,
WeeklyUsage.ts, WeeklySonnetUsage.ts, WeeklyOpusUsage.ts and
FableWeeklyUsage.ts. Normalized for the per-widget names, SessionUsage and
WeeklyUsage are identical, so are WeeklySonnetUsage and WeeklyOpusUsage,
and the worst pair differs by three lines. Net -225 lines.

Each file is now delegation, following the existing shared/speed-widget.tsx and
InputSpeed.ts pair.

Six axes vary and are parameterized: label, preview percent, usage data field,
window resolver, display name, description. Default color, category, editor
display, handleEditorAction, keybinds and the three supports* flags were
already uniform across all five, so they live in the shared module outright.
None of the five declares getHideableStates or renderEditor.

One rendered string changes. FableWeeklyUsage's label was the only one of
the five that disagreed with its own display name:

Widget	Display name	Label
Session Usage	Session Usage	Session:
Weekly Usage	Weekly Usage	Weekly:
Weekly Sonnet Usage	Weekly Sonnet Usage	Weekly Sonnet:
Weekly Opus Usage	Weekly Opus Usage	Weekly Opus:
Weekly Fable Usage	Weekly Fable Usage	Fable Weekly:

It becomes Weekly Fable: , matching the sibling rule of display name minus
Usage. Raw mode prints no label, so raw output is unchanged, as are the widget
type fable-weekly-usage, the class name and the display name. Six assertions in
FableWeeklyUsage.test.ts move with it; no other assertion in the repo changed.

Two things worth a look:

• The resolvers dispatch at call time, not through a captured reference. All
  five test files spy their resolver on the utils/usage namespace. A
  module-level config table holding resolveWeeklyOpusUsageWindow would bind at
  module evaluation, before beforeEach installs the spy, and silently bypass
  it. SessionUsage also needs a second argument (context.blockMetrics) that
  the four weekly resolvers ignore, so the parameter takes the whole context.
• ExtraUsageUtilization stays out. It gates on a tri-state
  extraUsageEnabled with its own n/a branch, has no window resolver, no time
  cursor, an extra keybind and a different default color. Folding it in would mean
  three optional flags that are uniformly absent from the five. It is also the one
  usage widget that does not use runUsagePercentWidgetSuite.

The shared suite gains a whole-percent case, which is the one test addition
beyond the label. formatPercent's format argument is optional and its default
reproduces the baseline output, so a render path that stopped passing the
resolved format rendered identically under default settings and nothing could
catch it. Dropping the argument at all three call sites now fails all five
widgets.

Tested: bun run lint clean; bun test 1900 pass, plus the
global-command-resolution failure main already has on this host and one hit of
the flaky fetchUsageData case.