Add Windows ROCm build instructions to HIP (ROCm) section

[0xDELUXA]

Summary

Adds a collapsed Windows ROCm subsection inside the existing HIP (ROCm) details block.

Why

The current HIP section only covers Linux. Windows users building with AMD's TheRock-based ROCm SDK (installed via pip into a venv) have no documented path. This adds one.

Verified working

Check build output with:

llama_cpp.llama_print_system_info().decode('utf-8')  # should contain "ROCm"
llama_cpp.llama_supports_gpu_offload()               # should return True

Output on test machine:

ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16304 MiB):
  Device 0: AMD Radeon RX 9060 XT, gfx1200 (0x1200), ...
ROCm : NO_VMM = 1 | ...
GPU Offload Supported: True

No existing behavior is affected. Linux ROCm users see no change. The new block is opt-in via click-to-expand.

[JamePeng]

OK, Thank you for providing the documentation, which helps improve the system and better assist more AMD accelerator card users. If you feel there's nothing else to add, I'll merge it later.

[0xDELUXA]

OK, Thank you for providing the documentation, which helps improve the system and better assist more AMD accelerator card users. If you feel there's nothing else to add, I'll merge it later.

Good to see you're open to adding this. I've maintained my fork for months now, it only had this single doc change. Haven't experienced compilation/stability issues, so I tweaked this line a bit as well. Now it's ready to be merged.

[0xDELUXA]

Thanks! If any Windows ROCm issues arise, feel free to ping me anytime.

[JamePeng]

OK :)