Add Windows ROCm 7.13 build support

[Crssz]

Overview

Edit cmake script to make ROCm 7.13 compile on windows with AMD GPU, I add win32 guard since my pc not available to install any linux distro.

Additional information

Tested on
Hardware: AMD Radeon AI PRO R9700 (gfx1201, RDNA4, 32 GB VRAM)
OS: Windows 11 25H2
Software: ROCm 7.13
Models: Qwen3.5-27B Q4_K_M

command

# no dflash
.\llama-server.exe -ngl 999 --port 3000 -c 8192 -m "F:\models\Qwen3.5-27B-Q4_K_M.gguf" -sm none -mg 1 -fa on --reasoning off
# dflash
.\llama-server.exe -ngl 999 --port 3000 -c 8192 -np 1 -sm none -mg 1 --reasoning off -fa on -m "F:\models\Qwen3.5-27B-Q4_K_M.gguf" -md "F:\models\dflash-draft-q4_k_m.gguf" -ngld 99 --draft-max 16 --spec-type dflash -np 1

prompt
Implement a word-guessing Hangman game in JS Canvas.

Results

Run	Prompt eval time (ms)	Prompt tokens	Prompt ms/token	Prompt tokens/sec	Eval time (ms)	Eval tokens	Eval ms/token	Eval tokens/sec	Total time (ms)	Total tokens
baseline	152.07	25	6.08	164.40	89028.54	2566	34.70	28.82	89180.61	2591
turbo4	155.46	25	6.22	160.81	78341.20	2139	36.63	27.30	78496.66	2164
dflash	161.16	25	6.45	155.13	61626.67	2967	20.77	48.14	61787.82	2992
dflash turbo4	166.71	25	6.67	149.96	47596.95	2332	20.41	48.99	47763.67	2357

Draft Acceptance

Run	Acceptance rate	Accepted	Generated
dflash	0.27292	2408	8823
dflash turbo4	0.29091	1897	6521

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES format test result from llama server log to markdown table