perf: keep chunk-K residency engaged with runtime LoRA

[fszontagh]

Summary

Re-enable chunk-K residency for the runtime LoRA path. Two related fixes in compute_streaming_segments / resolve_graph_cut_plan:

• Drop the weight_adapter != nullptr bypass. Runtime LoRA composes weight + diff in the compute graph via ggml_add; the resident weight is never mutated, so the cached GPU copy stays valid across sampling steps.
• Add the resident buffer size back to free_vram before clamping the streaming budget. Otherwise chunk-K's own allocation is read as "taken by someone else", the budget shrinks step-to-step, and the resident set rebuilds every step instead of every generation.

Related Issue / Discussion

Follow-up to #1576 (--stream-layers). Closes the LoRA-path perf gap left there.

Additional Information

Z-Image bf16, 512x512, 8 steps, --offload-to-cpu --stream-layers --max-vram 8 on RTX 3060:

Config	Before	After	Delta
LoRA 0.8	46.78 s	37.75 s	-19%
LoRA 1.5	45.83 s	37.14 s	-19%
no LoRA	44.64 s	34.72 s	-22%

LoRA multiplier scaling unchanged (0.8 vs 1.5 mean pixel diff 26.17 -> 26.05).

Checklist

• I have read and confirmed this PR follows the contribution guidelines.