Fix two latent Ground Zero bugs: random-respawn dmflag filters and tesla retargeting

[Niehztog]

Two latent bugs from the original Ground Zero source, both in code that
silently does the opposite of what it says. Neither shows up as a compiler
diagnostic, and the tree builds clean with and without the changes.

1. FindSubstituteItem tests the wrong item (src/g_newdm.c)

The DF_NO_SPHERES, DF_NO_NUKES and DF_NO_MINES filters test
ent->classname — the item being replaced — inside the loop that is choosing
the item to replace it with. The condition does not depend on the loop, so it
fires for every candidate or for none, and the effect is the inverse of the
comment above the sphere filter: a disabled sphere respawns as itself, while
every other item can still be substituted into one. SpawnItem's own
DF_NO_SPHERES guard then frees that substitute, and DoRandomRespawn returns
the freed edict to DoRespawn, which relinks it as a solid trigger.

The third sphere classname is also misspelled "item_spehre_defender", so of
the three the Defender is the one DF_NO_SPHERES never recognised.

The 2023 remaster tests the candidate item in all three filters and spells the
Defender correctly.

The sphere filter only existed in the counting pass. Once the filters depend on
the candidate, both passes have to apply the same ones or they disagree on how
many items are eligible and the pick walks off the count — so both now share one
predicate, which identifies spheres by Pickup_Sphere, the same test
SpawnItem already uses for this dmflag.

2. A monster already angry at a tesla is retargeted anyway (src/monster/misc/move.c)

SV_movestep originally had four cases: no valid enemy, retarget; the enemy is
already a tesla, do nothing; the enemy is a player, retarget only when it cannot
be seen; any other enemy, retarget. The second was guarded by
strcmp(ent->enemy->classname, "telsa") — a misspelling — so it never ran and
fell through to the last case, retargeting after all.

Collapsing the chain into a single condition preserved the behaviour of the
misspelling rather than of the case it was written for: a tesla is not a client,
so !ent->enemy->client is true for a monster whose enemy is a tesla and it is
retargeted on every blocked frame.

TargetTesla skips the switch when it is the same tesla, but its AI_MEDIC
bail runs before that test, so a medic repeatedly calls cleanupHealTarget on
a tesla; and when the blocking area belongs to a different tesla the monster
switches to that one and overwrites oldenemy with a tesla, losing the player
it was chasing.

The remaster keeps all four cases and spells the classname correctly, so this
restores the exclusion rather than reintroducing the typo.

Testing

make is clean with no new warnings, before and after, on gcc 13 (Linux
aarch64). The two changes are independent and in separate commits if you would
rather take only one.

[Yamagi]

Thank you :)