Avoid redundant memory_breakdown() computation in device memory probe

[Jessen-Li]

Overview

This PR removes a redundant computation of ctx->memory_breakdown() during device memory probing in llama_get_device_memory_data().

Additional information

Changes

• Introduce llama_memory_breakdown_print_impl which accepts a precomputed memory_breakdown object.
• Reuse the existing memory_breakdown result from llama_get_device_memory_data() instead of recomputing it in
  llama_memory_breakdown_print()

Motivation

llama_get_device_memory_data() may be called multiple times during fit/probe workflows (e.g. baseline, context probing, repeated resource checks). Each call currently recomputes memory_breakdown, leading to unnecessary duplicated work.

This change ensures:

• No redundant computation within a single probe
• Consistent snapshot used for both computation and logging
• Slight reduction in overhead in repeated probe scenarios

Behavior

No functional change is intended.
Output and memory accounting remain identical.

Requirements

No new requirements introduced by this change.

• I have read and agree with the contributing guidelines
• AI usage disclosure:
  NO.