Send the isolation mutex name on the wire's IsolationMutexName field

Summary

• Bug: HalibutMachineExecutionStrategy (both dispatch sites) built the agent isolation mutex name but passed it into StartScriptCommand's taskId parameter (arg 7) instead of isolationMutexName (arg 5), leaving the wire field null. The agent (ScriptIsolationMutex) then fell back to its single "default" mutex — so the per-action name the code + ARCH.7 docs + tests claimed to send never reached the wire, and the SHA hash polluted the correlation field. Linux/WindowsTentacleUpgradeStrategy passed null + the ticket id the same way.
• Decision (semantics): serialise FullIsolation scripts per machine — the safe default — made explicit via a shared ScriptIsolationMutexNames.ForMachine(machineId) used by both deployment dispatch and the upgrade strategies. So every FullIsolation script dispatched to a machine (deployment or upgrade) serialises behind one writer lock, instead of relying on the accidental null → "default" fallback. Crucially, deployments and upgrades now share the identical name, preserving the deploy⇄upgrade agent-level serialisation (they were both on "default" before; a one-sided change would have split them onto different mutexes).
• Removed the per-action GenerateMutexName (keyed by serverTaskId, which actually defeated cross-deployment serialisation — FullIsolation's whole purpose). taskId now carries the server task id for correlation; ResolveParentTraceId ignores it, so this is behaviour-neutral on the agent.
• Non-breaking: on a single-machine agent the global lock stays global — serialisation behaviour is unchanged. Only the wire field, the doc comments, and the tests now reflect it. No agent-side change required (it already reads IsolationMutexName, null-falling-back to "default"); old-server↔️new-agent and new-server↔️old-agent both still serialise.

Test plan

• Unit (red→green): new ExecuteScriptAsync_PassesMachineScopedMutexNameAsIsolationMutexName_NotTaskId captures the dispatched StartScriptCommand and asserts IsolationMutexName == ForMachine(machineId) + TaskId == ServerTaskId — the wire field no prior test pinned, which is why the mis-wiring survived (confirmed RED before the fix)
• Unit: ScriptIsolationMutexNamesTests pins the format literal + same-machine-same-name / different-machine-different-name (the shared-string invariant deploy⇄upgrade serialisation depends on)
• Unit: both upgrade strategies' isolation-pin tests now assert IsolationMutexName == ForMachine(machineId)
• Removed the stale per-action GenerateMutexName tests + ARCH.7 comments; corrected the ticket test to check ScriptTicket.TaskId (the actual Guid) not the SHA-in-TaskId
• 251 affected tests green; broad DeploymentExecution + Deployments + Machines.Upgrade sweep 5017/5017 green; dotnet build Squid.sln 0 errors
• Tentacle ScriptExecution suite: 30 LocalScriptServiceTests fail only on a near-full dev disk (DiskSpaceChecker 10% floor in StartScript — environmental, agent code untouched by this change, branch-independent; all 30 confirmed disk-space IOException, zero mutex-related; pass on CI)