P7: AllocationResolver — one resolver decides domain and scope #1143

Sub-issue of #1133. The P7 answer, implemented: placement is decided by a resolver — a pure function of (what the artifact holds × what the profile says × what the platform can do) — mirroring WeightFormResolver, the pattern that already works. Consumers carry and obey; they never decide.

Scope

• New sk.ainet.lang.memory.plan.AllocationResolver (@ExperimentalMemoryApi):
  resolve(form: WeightForm?, format: Format, elementCount: Long, bytes: Long, profile: PlannerProfile, platform: StorageCapabilities): AllocationSpec
  with a small injectable StorageCapabilities(supportsMappedFiles, …) whose current() reads PlatformStorage. Rules:
  • WeightResidency.MAPPED ∧ platform can map ∧ encoding kept-as-stored → MMAP_FILE / MODEL / immutable
  • MAPPED but unmappable platform, or re-encoded form (dequant output bytes are not file bytes) → profile.domainFor(residentBytes) + MODEL
  • non-weight / HEAP → profile.domainFor + caller's scope
• MemoryPlan.kt: replace the hardcoded PlanTensor.allocation property (unconditional MMAP_FILE/MODEL, ignores form.residency, profile, and platform) with fun allocation(profile, platform) delegating to the resolver. The property has zero main-source consumers, so the signature change is safe.
• This gives PlannerProfile.domainFor(bytes) its first real consumer.

Acceptance

• New AllocationResolverTest (commonTest) + extended MemoryPlanTest/PlannerProfileTest: MOBILE_2GB mapped Q4_K → MMAP_FILE; same form on a no-mmap platform → falls back via domainFor; small bias → HOST_HEAP; a dequantized form never yields MMAP_FILE; domainFor threshold edges.
• Full pr-gate green (pure common code, all legs).