scx_layered: add per-layer irq_protect with fallback policy

[hodgesds]

Layers can now opt in to clearing their CPUs from every IRQ's /proc/irq/<N>/smp_affinity mask via a new irq_protect: bool field on LayerCommon. This is useful for GPU training workloads where collective-communication threads (e.g. NCCL) need very low jitter and must not share CPUs with NIC/GPU IRQ handlers.

The protector snapshots every numeric IRQ's affinity at startup, applies fresh masks after each refresh_cpumasks based on the union of protected layers' CPUs, and restores originals on shutdown. Kernel managed IRQs that reject affinity writes are recorded after the first failure and never retried.

A new --irq-protect-fallback={spread,all} flag selects what happens when an IRQ's home affinity is fully covered by protected layers (or when every CPU on the system is protected):

• spread (default) preserves per-IRQ locality; per-IRQ spill to the unprotected set; and when no CPU is unprotected, pin each IRQ to a single CPU round-robin across all CPUs.
• all gives every IRQ the same mask: the system-wide unprotected set, or all CPUs when nothing is unprotected.

Mode transitions and per-IRQ spill events are logged once. Includes unit tests covering normal strip + spill, managed-IRQ skip, redundant write elision, restore, spread mode (basic + wrap), spread<->normal transition, and both All-mode behaviors.