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).

probe_paged_attn.py

@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+"""Probe ixformer.vllm_single_query_cached_kv_attention signature and test."""
+import inspect
+import torch
+import ixformer
+
+# Print signature
+fn = ixformer.vllm_single_query_cached_kv_attention
+print(f"Signature: {inspect.signature(fn)}")
+
+# Also check v2
+if hasattr(ixformer, 'vllm_single_query_cached_kv_attention_v2'):
+    fn2 = ixformer.vllm_single_query_cached_kv_attention_v2
+    print(f"V2 Signature: {inspect.signature(fn2)}")
+
+# Check contrib.vllm_flash_attn if available
+try:
+    from ixformer.contrib import vllm_flash_attn
+    print(f"\nvllm_flash_attn dir: {[x for x in dir(vllm_flash_attn) if not x.startswith('_')]}")
+except Exception as e:
+    print(f"\nvllm_flash_attn: {e}")
+
+# Check ixformer.vllm submodule
+try:
+    import ixformer.vllm as ixv
+    print(f"\nixformer.vllm dir: {[x for x in dir(ixv) if not x.startswith('_')]}")
+except Exception as e:
+    print(f"\nixformer.vllm: {e}")