fix: enable qwen36 GPU tier when building with HIP=1

[yuripourre]

Summary

• Fix make -C c qwen36 HIP=1 on Linux ROCm: the target linked backend_cuda.o but still used NOCUDA_LDFLAGS, which dropped -lstdc++ and failed at link time with DSO missing from command line.
• Mirror the existing CUDA=1 branch so HIP builds also compile qwen36_tier.c and link with the full CFLAGS/LDFLAGS set (including -DCOLI_CUDA, -lamdhip64, and -lstdc++).

Problem

The qwen36 Makefile switch only treated CUDA=1 as a GPU build. With HIP=1, $(CUDA_OBJ) (including backend_cuda.o) is still linked into qwen36, but the build fell through to the CPU-only branch:

• QWEN36_LDFLAGS = $(NOCUDA_LDFLAGS) — no -lstdc++ / HIP runtime flags
• QWEN36_TIER_SRC empty — GPU expert tier not compiled

Repro (Linux + ROCm):

make -C c qwen36 HIP=1 HIP_ARCH=gfx1201 HIPCC=/usr/bin/hipcc

Code changes assisted by Cursor.

[JustVugg]

Retargeted from main to dev. Every change lands on dev first; main only moves on a release, and it is protected, which is why the checks showed as blocked. Nothing to do on your side, the branch and the diff are unchanged.

[JustVugg]

The Linux failure here (expert did not become resident during warmstart, in test_qwen36_tier_multidev) is not yours. Your run was built against a dev snapshot taken in the middle of a merge sequence that landed #1390 and #1388, and #1388 changes assertions in that exact test. Both tests pass on settled dev (six consecutive runs of test_qwen36_tier_multidev, five of test_qwen36_tier_fill_wait). I have re-run the failed jobs against current dev. Nothing to change on your side; the four Makefile lines are the right shape.

[yuripourre]

@JustVugg thanks for clarifying, I will be more mindful in the next PRs.

Thank you so much for merging as well.