Add a multi-level feedback queue (MLFQ) scheduler

[mutehimself]

Summary

Closes #1096. Adds kernel/scheduler_mlfq, a new scheduler policy alongside the existing round-robin/priority/epoch schedulers, selectable via THESEUS_CONFIG=mlfq_scheduler.

• 8 priority levels; new tasks start at level 0 (highest priority), so short-lived/interactive tasks run promptly without ever being demoted.
• Quanta grow linearly with level depth, so tasks that prove themselves CPU-bound are scheduled less often but for longer stretches.
• Demotion is based on CPU time actually measured at dispatch, not on counting yields — a task can't dodge demotion by chunking CPU-bound work into pieces smaller than its quantum.
• A task that's no longer runnable when charged (it blocked rather than exhausting its quantum) is never demoted — this is the mechanism that makes MLFQ favor interactive/I/O-bound tasks over CPU-bound ones, with no static classification required.
• A periodic priority boost resets every task to level 0, bounding worst-case wait time and preventing starvation.
• Implements PriorityScheduler, so scheduler::inherit_priority (used by sync_block to prevent unbounded priority inversion) keeps working correctly under this policy — without it, a lock holder could sit at a low MLFQ level while higher-priority waiters starved behind it.

Also updates ps to show the priority/level column under mlfq_scheduler, matching its existing behavior for the epoch/priority schedulers.

Benchmark (addresses #758)

scheduler_eval's only mode measured aggregate time for N identical tasks to yield, which can't distinguish a scheduler that favors interactive tasks from one that doesn't, since every task behaves the same way. Added a -m/--mixed mode that spawns a configurable mix of CPU-bound tasks (busy loop, never yields) and interactive tasks (short work bursts, yields between each), and reports each group's completion-latency distribution separately.

Results, one run each, 4 CPU-bound + 16 interactive tasks pinned to one CPU (run headlessly in QEMU via a first_application swap + isa-debug-exit, same mechanism qemu_test uses):

Scheduler	CPU-bound avg latency	Interactive avg latency	Makespan
round-robin (default)	1382.3ms	772.0ms	1446.2ms
mlfq_scheduler	918.1ms	406.4ms	962.8ms

MLFQ roughly halves interactive-task latency relative to round-robin under identical CPU-bound contention, while also lowering CPU-bound latency and overall makespan (plausibly from fewer total context switches, since demoted tasks get longer quanta). Caveat: single run per scheduler under QEMU/TCG software emulation, so absolute numbers aren't representative of real hardware and the makespan difference in particular could partly reflect run-to-run emulation variance — the interactive-vs-CPU-bound ratio within each run is the more trustworthy comparison. Multiple trials and/or real hardware would firm this up further.

Test plan

• cargo check against the x86_64-unknown-theseus target for scheduler_mlfq, spawn/ps (with --cfg mlfq_scheduler), and scheduler_eval.
• Full make iso THESEUS_CONFIG=mlfq_scheduler build, booted in QEMU: clean SMP bring-up (4 CPUs), memory/ACPI/PCI/PS2/framebuffer init, shell loaded, no panics.
• scheduler_eval -m run headlessly under both mlfq_scheduler and default round-robin for the comparison above.
• Would appreciate maintainer input on whether BASE_QUANTUM/BOOST_INTERVAL (currently fixed constants) should be configurable, and on test coverage expectations for a new scheduler policy.