Fix llm chat cursor-key binding on Windows

[dchaudhari7177]

Refs #1639. I'm on Windows, so I could check this against a real pyreadline3 rather than reason about it — and what I found is a bit different from the issue's diagnosis, so the fix is shaped differently too.

What's actually happening

The Windows branch never did anything. It called:

readline.parse_and_bind("bind -x '\e[D: backward-char'")

pyreadline3 matches a binding line against \s*(\S+)\s*:\s*([-a-zA-Z]+)\s*$. bind -x '...' doesn't match, so it's logged as unparseable and dropped. On pyreadline3 3.5.4 the call returns without raising and binds nothing — so the arrow keys were never being configured on Windows, silently.

The GNU form isn't the answer either. pyreadline3 names keys left/right, not escape sequences:

>>> readline.parse_and_bind("\e[D: backward-char")
IndexError: Not a valid key: '\e[d'

And nothing needs binding there. pyreadline3's default keymap already has what the code was trying to add:

(False, False, False, 'left')  -> backward_char
(False, False, False, 'right') -> forward_char

The fix

Drop the Windows branch rather than correct it, with a comment recording why — otherwise the natural next move is to "fix" it to left: backward-char, which would be redundant with pyreadline3's own defaults.

Keep the POSIX bindings, wrapped. Binding cursor keys is a convenience; a readline shim that doesn't understand the syntax shouldn't stop llm chat from starting. That's the failure the issue reports, and the guard covers it regardless of which shim is involved — I couldn't reproduce a raise with pyreadline3 specifically, so I'd rather make the call non-fatal than guess at the reporter's environment.

Extracted into _bind_cursor_keys() so it's testable without invoking the whole chat command.

Tests

Three, all of which fail without the change:

FAILED TestBindCursorKeys::test_windows_binds_nothing
FAILED TestBindCursorKeys::test_posix_binds_the_arrow_keys
FAILED TestBindCursorKeys::test_a_readline_that_rejects_the_syntax_does_not_stop_chat

pytest tests/test_chat.py    5 passed, 8 xpassed
ruff check llm tests         All checks passed!
black --check                2 files would be left unchanged

Incidentally: tests/test_chat.py is decorated @pytest.mark.xfail(sys.platform == "win32"), and 8 of those xpass on this machine. Not touching that here, but it may be worth revisiting whether the Windows xfail is still earned.