[addon-operator] fix dedupe one startup

[ldmonster]

Overview

PushRunModuleTask is supposed to skip pushing a moduleRun task when the main queue already holds one for the same module. The guard behind it only answered true for pending tasks carrying DoModuleStartup, so it never fired for the way callers actually push — with doModuleStartup=false. This PR makes the guard consider the flag of the push itself, and fixes the pending-task scan that feeds it.

What this PR does / why we need it

Two related fixes in pkg/module_manager/module_manager.go:

1. The dedup guard ignored non-startup pushes. queueHasPendingModuleRunTaskWithStartup(q, moduleName) returned meta.doStartup, i.e. it only reported a duplicate when the pending task ran the startup sequence. Every caller pushes with doModuleStartup=false, so the guard returned false for all of them and PushRunModuleTask appended a task every time it was called. The main queue filled up with identical moduleRun tasks and the module was rerun — and re-released — once per duplicate.

   It is now queueHasPendingModuleRunTask(q, moduleName, doModuleStartup) and answers whether a pending task covers this push:

   pending task	push	covered
   with startup	with startup	yes
   with startup	without startup	yes
   without startup	without startup	yes (the fixed case)
   without startup	with startup	no

   The last row is the asymmetry worth keeping: a pending task that does strictly less work must not swallow a push that needs the startup, otherwise the startup is silently dropped.

2. modulesWithPendingTasks let the last task seen decide for the module. The scan assigned modules[hm.ModuleName] = struct{...}{doStartup: hm.DoModuleStartup} per task, so when a module had several pending tasks, a later one without the flag overwrote an earlier one with it — and a queued startup looked absent. It now ORs the flag across all pending tasks for the module.