Improve test coverage for balancing up to weeks (when year/month)

[arshaw]

I noticed fullcalendar's temporal-polyfill has an incorrect implementation of balancing days and smaller up to weeks when years or months are present. It was not getting caught by test262 tests. We should improve test coverage for this.

Temporal.Duration.from({ months: 1, days: 1 })
  .round({
    relativeTo: '2024-01-01',
    smallestUnit: 'weeks',
    roundingIncrement: 99,
    roundingMode: 'ceil'
  })
  .toString()

// temporal-polyfill: P2M (wrong)
// reference impl:    P1M99W (correct)

Things work fine in the reference implementation.