fix: LRU(2) eviction for prepack cache — prevents OOM across 61 layers

Cache was growing unbounded — 61 MoE layers × 2 = 122 prepacked SFB
tensors permanently in GPU memory (~1.75 GiB each). With sequential
layer execution, only 2 entries are needed at a time (current L1 + L2).
Added LRU eviction to keep max 2 entries.

src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py

@@ -106,7 +106,11 @@ def _prepack_weight_sf(weight_sf, N, K, tag):
     )
     if not hasattr(_prepack_weight_sf, '_cache'):
         _prepack_weight_sf._cache = {}
+        _prepack_weight_sf._cache_order = []  # LRU order
     if cache_key in _prepack_weight_sf._cache:
+        # Move to end (most recently used)
+        _prepack_weight_sf._cache_order.remove(cache_key)
+        _prepack_weight_sf._cache_order.append(cache_key)
         return _prepack_weight_sf._cache[cache_key]
 
     assert weight_sf.dtype == torch.float8_e4m3fn, weight_sf.dtype
@@ -127,6 +131,12 @@ def _prepack_weight_sf(weight_sf, N, K, tag):
 
     packed = torch.stack(packed, dim=0).contiguous()
     _prepack_weight_sf._cache[cache_key] = packed
+    _prepack_weight_sf._cache_order.append(cache_key)
+
+    # Evict oldest entries — keep only 2 (current layer's L1 + L2)
+    while len(_prepack_weight_sf._cache) > 2:
+        oldest = _prepack_weight_sf._cache_order.pop(0)
+        del _prepack_weight_sf._cache[oldest]
 
     if MEGA_MOE_DEBUG:
         print(f"[PREPACK] {tag}: E={E} N={N} K={K} packed_shape={packed.shape} "