perf(ROCm): optimize chunked CE loss num_warps for RDNA GPUs (#4123)

Use 16 warps for RDNA in the chunked cross-entropy forward kernel
(large vocab > 65536), matching the existing CDNA optimization.

Benchmarked on W7900 (gfx1100) with actual unsloth kernels (5 trials, median):
  - Chunked CE forward (BS=65536): 16 warps = 2.4-2.6x faster than 32
  - All other kernels (LayerNorm, RoPE, SwiGLU): default heuristic is
    already optimal for RDNA; no modification needed.

Depends on: #4109 (provides is_rdna() detection)

unsloth/kernels/cross_entropy_loss.py

@@ -22,6 +22,7 @@ from .utils import (
     triton_cast,
     torch_gpu_device,
     is_cdna,
+    is_rdna,
 )
 from transformers.models.llama.modeling_llama import logger
 from unsloth_zoo.utils import Version
@@ -364,7 +365,7 @@ class Fast_CrossEntropyLoss(torch.autograd.Function):
                     SOFTCAP = logit_softcapping,
                     DO_LOGIT_SCALING = DO_LOGIT_SCALING,
                     LOGIT_SCALE = logit_scaling,
-                    num_warps = 32 if not is_cdna() else 16,
+                    num_warps = 16 if is_cdna() or is_rdna() else 32,
                 )
             # logsumexp(chunked_logsumexp) - x
             # Do the -x separately