refactor policies to per-process scheduling granularity

[kylin1019]

Summary

Refactor UP, CFS, MLFQ, KEDF, and LAX policies from per-XQueue to per-process scheduling; add SwitchProcess() to suspend/resume all XQueues of a process as a unit; trigger immediate Sched()+ExecuteOperations() on policy switch in LocalScheduler::SetPolicy().

Motivation

LLM serving frameworks (llama.cpp, vLLM) create multiple CUDA streams per process. Per-XQueue policies can suspend one stream while another stream of the same process is running, causing a circular wait: the running stream blocks on an inter-stream dependency that requires the suspended stream to resume, but the scheduler waits for the running stream to idle before rescheduling. Per-process scheduling eliminates this deadlock by treating all XQueues of a process as an indivisible unit.

Scope

Only scheduler policy refactoring and the SetPolicy() fix. No new policies, hint types, HTTP API changes, tests, or docs.

Validation

• Builds with -Wall -Wextra -Werror;
• Tested with llama.cpp dual-instance serving: KEDF serializes correctly, CFS respects weights, UP enforces utilization direction;
• No deadlock observed.