[theAndreas]

Fixes what a review turned up, all of it carried into the TCA0 port in
AndreasBur/Wordclock#79 as well, where the same
fixes are in the third commit.

The faults

• read() divided before it shifted, so the prescaler scaled the truncation up. At 1024 it moved
  in steps of 1024 us where a tick is 32 us: a counter of 31 read as 0 us instead of 992. Shifted
  first, it is good to one tick, and 2 * 0xFFFF << 10 still fits in 32 bit.
• setPeriod() wrote PER outright. In dual-slope mode a TOP below the counter mid-period is
  not met until the counter has gone round 0xFFFF, one period of up to four seconds in the wrong
  place. It now goes through PERBUF, taken over at BOTTOM.
• The duty did not follow the period. It became an absolute compare value once, so
  init(1000), 50 % and then setPeriod(2000) ran at 25 %. The duty is now kept per channel and
  the CCxBUF are recomputed with PERBUF, so both arrive at the same BOTTOM.
• The first period after start() ran on CCx = 0, the value init()'s RESET leaves, because
  the buffers arrive only at the first BOTTOM. start() now loads PER and CCx directly while the
  timer has no clock. A forced UPDATE before the RESTART would not do: both commands go to
  CTRLGSET, and an UPDATE still pending there, ORed with RESTART, reads as RESET.

Smaller fixes:

• setPeriod()'s "set to maximum" branch could not be reached. An out of range period, and 0 us,
  is now refused outright.
• attachInterrupt() clears an overflow flag left over from while the interrupt was off.
  Otherwise the new callback ran at once.
• stop() does nothing before init(), when the timer may still belong to the core.

Configuration

TIMER_USE_TCC4 and TIMER_USE_TCC5 can now be set with -D. Selecting a timer that is switched off is a
static_assert instead of an undefined reference at the link. F_CPU has to be a multiple of
2 MHz, also asserted, because setPeriod() and read() work in whole cycles per microsecond.

What was checked

• Built with avr-gcc 14.3 for atxmega32e5 at -Wall -Wextra, with TCC5 alone and with TCC4
  switched on.
• Both assertions made to fail on purpose: Timer<TimerC4> with TCC4 off, and F_CPU at 12.5 MHz.
• The read() error was worked through for prescalers 64 and 1024.
• 192 bytes of flash and 6 bytes of RAM more for a program that uses the whole API on TCC5
  (1212 -> 1404 bytes, Timer.cpp linked at -Os -flto against an empty main()). The RAM is
  Top and one duty per channel.

Not run on hardware.