perf(loongarch64): bypass event queue for virtual ipi

[li041]

Summary

• Update the virtual IPI pending-status bitmap before notifying the target CPU.
• Use a dedicated physical IPI doorbell (bit 4) for virtual IPI delivery instead of allocating a generic event.
• Preserve generic event-queue processing for other queued work by draining it only when events are pending.
  • The generic event-queue may be removed by a explicit vcpu abstraction. See Tracking issue: Introduce an Explicit vCPU Abstraction #395 for details.

Close #383

Motivation

Virtual IPI delivery previously allocated a generic event solely to trigger virtual IPI line synchronization. This added unnecessary event-queue traffic and coupled posted virtual IPI state to generic event handling.

This change treats virtual IPI status as posted state: the status bitmap is updated first, and the target CPU is notified through the dedicated doorbell. Generic events remain available for wakeups and other deferred work. However, from a design perspective, the generic event handling mechanism incurs non-negligible overhead, and really requires a refactor.

Verification

• Hardware validation on the LoongArch 3A6000 board
• cyclictest with hackbench backend

Risks and Limitations

• The dedicated virtual IPI doorbell and related interrupt handling are currently specific to the LoongArch64 implementation.
• Generic events still use the per-CPU FIFO and retain its synchronization and processing overhead; this change only removes virtual IPI notifications from that path.

[agicy]

LGTM.