fix: keep echo on in the cooked and well-done modes

[Marukome0743]

Fixes #49.

What was wrong

vt_cooked() and vt_well_done() go through write_termios, which reads the
current attributes, wipes every flag field, and re-inserts a hardcoded set. That
set never contained ECHO, so after either mode runs the terminal accepts input
without showing it. VtWellDone is what watchexec --clear=reset applies, which
is how this reaches users — the symptom is reported downstream at
mise#8269.

Neither of the two commands people reach for fully undoes it: stty sane restores
the echo flags but leaves ISTRIP set, and reset clears ISTRIP but leaves
IEXTEN off. Issue #49 has the measurements.

What this changes

• Set the echo flags. ECHO, ECHOE, ECHOK, and ECHOCTL/ECHOKE on the
  platforms where nix exposes them, plus IEXTEN. Both modes share one
  usable_local_flags() so they cannot drift apart.
• Stop setting ISTRIP. It clears the eighth bit of every input byte, so no
  multi-byte UTF-8 sequence can survive to be interpreted as UTF-8 — which
  contradicted the IUTF8 that vt_well_done sets beside it.
• Carry the character size and the line speed across the reset. They describe
  the line rather than how the terminal behaves on it, and a screen clear has no
  way to know what to put back. The wipe left CSIZE at CS5 and, on Linux where
  the speed lives in the control flags, the baud rate at B0. A pty absorbs most
  of that, but on a real serial console B0 means hang up the line.

The rest of each mode is unchanged, and the doc comments are updated to match —
they enumerate the flags, so they were describing the old behaviour.

I went with keeping the authoritative from-scratch write rather than switching to
read-and-modify, since that design is deliberate and documented. If you would
rather have the other approach, I am happy to redo it.

Structure

The two mode bodies move from closures into named cooked_mode / well_done_mode
functions, and the duplicated body of write_termios moves into a
write_termios_to(fd, f) that takes the descriptor. That removes the copy-pasted
tty//dev/tty branches and lets the tests drive the real mode functions against a
descriptor of their own.

Tests

Four unit tests, each run against both modes, on a pty from nix::pty::openpty so
they never touch the test runner's own terminal:

• echo_survives — the echo flags, IEXTEN, ICANON and ISIG are all set afterwards
• the_eighth_bit_is_not_stripped — ISTRIP is not set afterwards
• the_line_configuration_is_left_alone — CSIZE and the speed come out as they went in
• a_terminal_already_without_echo_is_recovered — starting from a terminal with every
  local flag cleared, which is what these modes exist to rescue, still yields echo

nix::pty is behind the term feature, which is already enabled, so there is no
dependency change. The test module is cfg'd off the targets where openpty or
nix::pty do not exist (redox, fuchsia, aix), matching the crate's existing
target handling.

What I could not verify locally

I have no Rust toolchain on this machine, so I have not compiled this — I checked
every API against the nix 0.31.3 sources (tcgetattr/tcsetattr are generic over
AsFd, openpty is term-gated and absent on aix, ECHOCTL/ECHOKE are absent
on redox, libc_bitflags! derives Debug), and formatted with rustfmt against the
repo's .rustfmt.toml. CI is the real check here.

[passcod]

Co-Authored-By on the commit please :)

[Marukome0743]

Co-Authored-By on the commit please :)

Thanks for the quick review!

Added a Co-authored-by: Claude Opus 5 <noreply@anthropic.com> trailer and force-pushed.
The code itself is unchanged — commit message only.