feat(llama): multi-GPU layer distribution + KV cache placement

[MegalithOfficial]

Summary

Adds full multi-GPU support to the llama.cpp local runtime: users can split a model
across multiple discrete GPUs with a choice of distribution strategy, control where the
KV cache lives, and (per model) assign exact per-GPU and CPU layer counts. Builds on the
initial multi-GPU pass (db3deaf) and makes it actually configurable.

What's included

Distribution strategies (the "offloader mode"):

• Balanced — even layer split across the selected GPUs
• Proportional to VRAM — weighted by each GPU's free memory (good for mismatched cards)
• Priority fill — fill the first GPU up to a VRAM limit, then overflow to the next
• Manual per-GPU — explicit layer-count input per GPU plus a CPU layers input, with
  live GPU + CPU = total layers validation (per-model only)

KV cache placement: Auto / Split with layers / System RAM / Pin to one GPU
(maps to offload_kqv + main_gpu).

Guardrails:

• Integrated GPUs are never used — excluded from enumeration, selection, validation, and
  VRAM accounting (scoped to the multi-GPU paths; single-GPU iGPU-only hosts are unaffected)
• The Multi-GPU toggle is disabled when fewer than 2 discrete GPUs are present, on both the
  per-model editor and the global runtime-defaults page
• Only layer split is supported — row/tensor split were removed; they require a fast
  interconnect (NVLink) that's rare, while layer split is correct for PCIe consumer GPUs

Where it lives:

• Per-model: EditModelPage gets the full controls (strategy, devices, manual counts, KV)
• Global: LocalRuntimeDefaultsPage sets the inheritable defaults (strategy, priority limit,
  KV placement); per-model values fall back to these via the session → model → settings chain

i18n: the multi-GPU runtime-defaults strings (previously English-only) are now translated
across all 19 non-English locales.

Implementation notes

• No llama-cpp-rs fork change needed. The loader builds the raw llama_model_params
  struct, so tensor_split (per-device proportions) and main_gpu are set directly. The
  tensor_split buffer is held at function scope so it outlives llama_load_model_from_file.
• New plan_multi_gpu_distribution (offload.rs) turns a strategy + per-device free VRAM +
  model layer count into a concrete (n_gpu_layers, tensor_split, main_gpu).
• Manual mode pins a fixed total and bypasses the smart backoff ladder; auto strategies keep
  the ladder (proportions stay fixed while the total shrinks on VRAM-pressure fallback).
• model_params_key includes the new fields so a strategy change forces a reload.

Testing

• bun run check (tsc + cargo check) passes clean, no warnings.
• Verified locally: single-GPU regression (controls hidden, behavior unchanged), toggle
  disabled with <2 discrete GPUs, manual-sum validation, context-info estimates, and all 19
  locales (25 keys each, placeholders intact).
• ⚠️ Actual cross-GPU split, priority-fill overflow, and KV pin/system-RAM placement need QA
  on real 2+ discrete-GPU hardware — the dev machine has a single discrete GPU.