Rework the Windows timezone implementation

[dkhalanskyjb]

• Improved memory usage during timezone database initialization: now, buffer allocations are performed outside of loops.
• Removed the last production-facing piece of cinterop. cinterop is still used in tests.
• Added a thorough test to check that every timezone transition that we discovered is also a valid transition from the point of view of Windows.
• Used the test to determine the logic that Windows uses for time zone handling, fixing all inconsistencies but a couple of small ones:
  • In some cases, Windows claims that the transition happens on 23:59:59.999, whereas we can express 24:00:00.000, the value that was actually intended.
  • For a couple of time zones, for years 2030+, Windows seems to use some custom logic that doesn't seem correct anyway, so we don't reimplement it.