Clear the region before snapping point to the live viewport - #701

Fixes #698.

The bug

Switching to a live input mode snaps point to the end of the buffer. Three
call sites do this, and none of them cleared an active region first, so the
selection stretched from its start to the prompt. Emacs then published that
grown span to the primary selection when the mark was finally deactivated,
via select-active-regions. The kill ring is untouched and the grown region
is never highlighted, which is why this stayed invisible.

The reporter hit it after leaving the buffer and returning, which silently
enters a read-only mode by two independent routes: the minibuffer-exit check
sees point off the live cursor, and a buffer switch finalizes the mark
activation. The next keystroke exits that mode and triggers the snap.

This is not specific to ghostel-mouse-drag-input-mode being nil. Under the
default settings a mouse drag already enters copy mode, so drag-then-type
reproduces it with a stock configuration and no buffer switch.

The fix

deactivate-mark immediately before the snap, at each of the three sites:
the read-only exit, ghostel-semi-char-mode, and ghostel-char-mode. The
last two are the shared root for every entry direction, including entering
char mode from semi-char and leaving line mode.

The custom exit branch is left alone. It keeps point where the user put it,
and the one implementation that exists deactivates the mark itself.

Verification

Reproduced live in a sandboxed Emacs running a real shell under ghostel,
both reported routes and the default-configuration route, before and after.

Two ERT tests record point from deactivate-mark-hook, which Emacs runs
from inside deactivate-mark after clearing mark-active, so a deactivation
that happened after the snap reports the wrong position. Both were confirmed
to fail against the unfixed source with the compiled files rebuilt between
runs.

make -j8 all passes.