fix: read the redirect log as bytes, not text (CII-240)

[hacker-cb]

Fixes #425.

_PopenRedirectProcess._forward_io read the redirect log in text mode while it was still
being appended to. Every read reaches a temporary EOF, which finalizes the incremental
decoder, so a UTF-8 sequence straddling that boundary raises UnicodeDecodeError. The
surrounding except Exception swallows it — the process exits 0 with an empty stderr and
DUT output stops for the rest of the session, surfacing later as a pexpect timeout that
names no cause.

Why binary rather than a decoder

The decode was never needed here:

• MessageQueue.put accepts str | bytes and re-encodes through to_bytes(), which is
  the identity for bytes;
• PexpectProcess's buffer is bytes;
• the serial transport already forwards read_all() undecoded
  (pytest-embedded-serial/pytest_embedded_serial/serial.py:244).

So today's path is a bytes → str → bytes round trip. Reading in binary removes the failure
mode and makes the two transports consistent.

One behaviour change

Text mode applied universal-newline translation, so a DUT sending CRLF was forwarded as LF.
Binary mode preserves CR — which is what the serial transport already delivers, and both
_listen and the Unity parser handle CRLF already. test_forward_io_preserves_carriage_returns
pins it so it stays deliberate.

Tests

Two unit tests next to the existing test_listen_no_data_loss_*, in the same shape.

The reproduction condition is worth noting: the multi-byte sequence has to be split
between two appends, which a byte-at-a-time UART/QEMU stream does naturally. Writing a
whole line in a single write() does not reproduce it.

On unmodified main both fail:

AssertionError: forwarder died on a split sequence (exitcode=0)
AssertionError: assert b'first\nsecond\n' == b'first\r\nsecond\r\n'

With this change both pass.

Test runs on this machine

• pytest-embedded/tests/ — 36 passed, 1 skipped, 2 failed.
• The other eight packages — 47 failed, 24 passed, 9 skipped.

All of those failures reproduce on unmodified main: I ran the same selection on both refs
and the sorted FAILED lists are identical (empty diff), with the same 47/24/9 totals.
They need QEMU binaries, serial ports, an Arduino CLI or NuttX images, none of which are
present here. test_temp_disable_packages fails for a separate reason — it expects
ImportError for the sibling packages, which the editable install from CONTRIBUTING makes
importable.

Not included

Narrowing except Exception. MessageQueue.put already swallows a closed queue itself, and
the only escape I observed at teardown was BrokenPipeError, but I have not shown that list
to be exhaustive — so this PR stays one change. Happy to add it here if you would rather have
it together.

[github-actions]

Title	Coverage	Tests	Skipped	Failures	Errors	Time
Qemu		7	0 💤	0 ❌	0 🔥	25.154s ⏱️
3.14 X64		119	16 💤	0 ❌	0 🔥	15m 37s ⏱️
3.10 ARM64		119	18 💤	0 ❌	0 🔥	12m 46s ⏱️

[hfudev]

LGTM. Thank you for the fix. I'll do the release next Monday