Fix evil twin, pheasant and random event graphic bugs

[HarleyGilpin]

Closes #1227. Replaces #1233, whose branch picked up duplicate commits from an upstream rebase.

Evil Twin

With world.npcs.randomWalk on, the npc tick re-applies Wander over EmptyMode every tick, so a grabbed suspect strolled around mid-lift and wandered back out of the jail while the face-north flag was applied. WorldTest forces randomWalk off, which is why no test caught it. The grab, the drop at the jail and the re-entry spawn now use PauseMode, and the twin turns north on the same tick as the jailed animation, a tick after clearTransform, so the face update can't be lost to the transform or the walk's leftover facing.

New test runs the grab with randomWalk enabled and asserts the twin ends on the jail tile facing north.

Freaky Forester

• Pheasants die to a single hit from any attack style, even when the accuracy roll misses, handled in Target.damageLimitModifiers alongside damage_cap and immune_death.
• They also can't deal damage. Damage.maximum reads max_hit_melee off the npc definition ahead of the combat definition's range, so the inherited max_hit_melee = 10 had to go for the combat definition's target_hit of 0 to apply.
• The defend anim was wrong because pheasants used the chicken combat definition, and the chicken anims (5387-5389) belong to the reworked chicken rig the pheasant model never got. Pheasants now have their own combat definition with the old bird rig anims (2371-2373) and pheasant sounds (690-692).

New test forces every accuracy roll to fail and asserts one swing kills the pheasant and drops the raw bird; it fails without the fix.

Random event puff of smoke

Adds a shared eventHerald for the spawn and poof for the removal, and routes every event through them. poof spawns the graphic on the npc's tile rather than flagging it on the npc, the way Imp already does, so removal can't swallow it.

New test asserts the graphic is still flagged on the tick the npc reaches clients, since the spawn is queued a tick ahead of that.

Area exit handlers on the way into an event

Reported after the above: being taken by a random event out of Mort Myre kept the swamp decay running, and out of the barrows crypts kept the barrows overlay on screen for the whole event.

Movement.move subtracted the player's instance_offset from both ends of a move. The six events that build an instance set that offset before the teleport that carries the player in, so the from tile - still real map - was offset into a tile the player was never on, and no area they actually left fired its exit handler. Each end is now mapped on its own, and only when it is inside the block of regions instances are allocated from.

Fight cave was relying on that asymmetry. Its exit tile sits inside tzhaar_fight_cave_multi_area, so leaving on foot is not an area change and never was one; the instance end just used to be compared against a meaningless tile. It now tears the instance down in leave, and the exit handler keeps the logout path it always had.

Swamp decay had no exit handler at all and checked immunity ahead of the area, so a player carrying a blessed sickle out of the swamp kept the timer, the message and the shield graphic for the rest of their session however they left.

New tests kidnap a player out of the barrows crypts and out of Mort Myre and assert the overlay closes and the timer stops; both fail without the fix.

Fight cave progress lost on shutdown

World.shutdown despawned every player without marking them as leaving the world, so every area exit handler took its walked-out branch. Fight cave cleared fight_cave_wave and left the player standing on an instance tile, and the shutdown save persisted both - a restart mid-cave cost the run and logged the player back in inside a dead instance. Godwars killcounts and greegrees were dropped the same way.

logged_out is now set before the despawn, which is the flag all of those handlers already check. Logging out normally was never affected; AccountManager.logout sets it.

New test takes a player into the cave, shuts the world down and asserts the wave survives and the player is back on real map; it fails without the fix.

Flippa and Tilt chatheads

Both pinball trolls had no dialogue key, so NPCDialogue fell through to the default expression names. Those are the reworked human chathead set (expression_neutral1 = 9807 onwards, skeleton 2165); Flippa and Tilt share chathead model 3790, a 2005 troll head rigged to the pre-rework skeleton 82, so the frames tore it apart mid-dialogue.

They now use the old family (554-617, skeleton 82), which is the same set void already gives other pre-rework models - frogs, clay golems, etc.

Closes #1113

Pinball posts

The next post was drawn from all five, so it could land on the post that had just been tagged - which reads as the event having ignored the tag rather than as a new target. The post that was just lit is now excluded from the draw.

Tagging was also silent. The 634 cache carries this event's audio: 2278 is the pillar tag and 2279-2283 are one tone per pillar.

Tests now tag whichever post is flashing rather than assuming it never moves, plus a new case asserting the target never repeats across eight tags.

Evil Bob fishing spot pan

Sped up the camera pane and fixed a camera lock bug caused by moving after the last tick of the camera pane cutscene.