server: add margin for draft model for fit

[am17an]

Overview

Use common_get_device_memory_data to estimate fit params for draft model, and add that as margin before calling the fit on the main model. Resolves #23472

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, my local Qwen helped a bit.

[bartowski1182]

maybe out of scope, but would it be worth tracking the original fit_params_target and if loading MTP goes over the GPU threshold, but not loading it wouldn't have, print a warning?

[miloslavnosek]

Hopefully I won't clutter this PR, but does any of this affect the scenario where the draft model is loaded with spec-draft-model = xx.gguf rather than being baked in?

[sourenaraya]

Tested this branch on my 5070ti+7900xt setup (vulkan), looks like fit does not quite work as expected (at least — expected by me).
Loaded Qwen3.6-27B-UD-Q4_K_XL, VRAM usage ended up something like this:
5070Ti: ~13.4Gb/15.9
7900XT: ~19.6/19.9 (maxed out, basically).

What's interesting, VRAM usage increased in steps:

1. 7+9.5
2. 13.4+16.2
3. 13.4+19.6

After load any request to the model triggers OOM on amdgpu. Please note, that i set amdgpu.gttsize=32, because llama.cpp has very annoying tendency to use GTT and tank performance.
I end up tuning tensor-split manually to not spill to GTT.
fit.log