perf: native ixformer decode (v1 ≤32K, v2 >32K) + flash_attn_varlen prefill

Replaces all Python PyTorch fallback attention with native ixformer kernels:

Decode path:
- ≤32K: paged_attention_v1 (5D KV layout, x=8) — verified on real BI-V100
- >32K: paged_attention_v2 (5D→4D permute) — verified 65K+ on real BI-V100
- Removes _forward_decode_pytorch Python fallback entirely

Prefill path (profiling):
- _run_sdpa_fallback now uses ixformer.flash_attn_varlen_func
- head_dim=256 verified correct (diff<0.004) and 1.7x faster than PyTorch
- Falls back to Q-tiling pure-math if ixformer unavailable

Also includes: MoE kernel integration, GDN C++ kernels, diagnostic scripts,
xllm upstream layer/kernel references, .dockerignore cleanup.

All changes verified on real BI-V100 hardware (single card).

computility-run.yaml

@@ -8,14 +8,14 @@ command:
     - --served-model-name
     - llm
     - --max-model-len
-    - '131072'
+    - '80000'
     - --gpu-memory-utilization
     - '0.90'
     - --trust-remote-code
     - -tp
     - '4'
     - --max-num-seqs
-    - '1'
+    - '2'
     - --disable-log-requests
     - --disable-frontend-multiprocessing
     - --max-num-batched-tokens
@@ -46,4 +46,6 @@ env:
     - name: BI100_GDN_RESTORE_MODE
       value: hybrid64
     - name: BI100_MOE_COREX_TOPK_SOFTMAX
-      value: '0'
+      value: '1'
+    - name: PYTORCH_CUDA_ALLOC_CONF
+      value: expandable_segments:True