feat: Task priority levels for background tasks (interactive vs batch)

[alexeymorozua]

Problem

All background tasks currently have equal priority. When a user sends a message while batch tasks are running (e.g. a long music generation batch or a heavy analysis task), the interactive response competes with background work for the same resources.

This is particularly noticeable with long-running agent tasks (5+ hours) where the user's direct messages should always be processed first.

Proposed Solution

Add a --priority flag to task creation with MLFQ-inspired levels (reference: AgentRM paper, arxiv 2603.13110):

# Interactive (highest priority — user-facing)
create_task.py --priority interactive --name "..." "..."

# Background (default — normal async work)
create_task.py --priority background --name "..." "..."

# Batch (lowest priority — long-running bulk work)
create_task.py --priority batch --name "sonic-batch" "..."

Priority Levels

Level	Use case	Preemptable?
interactive (0)	Direct user request, needs fast response	No
background (1)	Standard async task	Yes, by interactive
batch (2)	Long-running bulk work (music gen, large analysis)	Yes, by higher levels

Anti-starvation

Lower-priority tasks that have been waiting too long should be promoted to prevent indefinite starvation — similar to Linux CFS nice values.

Impact

• User messages always feel responsive even under heavy background load
• Explicit intent declaration at task creation (no guessing)
• Safer long-running batch work without blocking interactive use

Note: Current LockPool handles per-session locking well; this is complementary — about scheduling order, not mutual exclusion.