hd44780: fix state corruption during 4-bit read-back (busy-flag polling)

[felixmertins]

Fixes #582

Two independent defects made the HD44780 part unusable with firmware
that polls the busy flag in 4-bit mode (standard bidirectional wiring as
in examples/board_hd44780). Details, cycle traces and analysis in the
linked issue; summary:

1. hd44780: don't reinterpret our own readback as pin input —
   hd44780_process_read() raises IRQ_HD44780_ALL with the nibble it
   drives back to the AVR; the part's own pin hook decoded that value as
   a full pin set and clobbered RS/RW/E in pinstate, so the discard
   strobe of a busy poll was processed as a garbage write. The ALL
   case now honours HD44780_PRIV_FLAG_REENTRANT, exactly like the
   D0 ... D7 case already does.

2. hd44780: track the data pins while driving a read back — the
   D0 ... D7 early-return under REENTRANT left pinstate stale
   while the FILTERED ioport pin IRQs cached the read-back level; a
   subsequent firmware write of the same logic level produced no edge
   notification and was latched with old data bits (single-bit
   corruption in characters written after busy polls). pinstate now
   tracks the value the part drives itself — it is the real bus level,
   nothing in the read path samples it, and the FILTERED D IRQs keep
   the bidirectional connection loop-free.

Validation: my AVR project uses an HD44780 driver that busy-polls before
every byte. Unpatched, the display stays empty / fills with garbage
(traces in the issue). With these two commits, the full LCD regression
suite (init sequence, both DDRAM lines, writes right after busy polls)
passes with no board-side workarounds. No change for firmwares that use
timed writes only, and external users of IRQ_HD44780_ALL as an input
are unaffected.

[gatk555]

It looks like LLM outout. What shows it is correct?

[felixmertins]

Yes, I use an LLM when writing things up. The findings themselves come from debugging my own firmware against this part, but fair point, prose doesn't prove anything. So I've added a regression test to the branch instead.

The firmware (tests/atmega644_hd44780.c) brings the LCD up in 4-bit mode and polls the busy flag before every byte, wired bidirectionally like examples/board_hd44780/charlcd.c. The test driver then compares DDRAM against what was written. Results:

• master without these two commits: the init sequence derails, DDRAM stays empty, test fails
• with only the first commit: single bits are dropped from characters written after busy polls, DDRAM line 2 reads "SO......ET" instead of "SOFT RESET", test fails. That is the corruption pattern from the issue.
• with both commits: passes, and the other 21 tests in the suite still pass

The mechanism is also easy to check by hand on master: hd44780_process_read() raises IRQ_HD44780_ALL with the read-back nibble (hd44780.c line 313), and the ALL case of the part's own pin hook (line 370) decodes RS, RW and E out of that value, without the REENTRANT guard that the D0..D7 case right below it has. So every status read rewrites the part's idea of RW.

The test hooks into the existing tests/ machinery and runs as part of make run_tests. It needs one extra make rule because it links examples/parts/hd44780.c, which is not part of libsimavr.