[Misc] Enhancements of shared memory layout and cluster api

[yaoyaoding]

1. Shared Layout Enhancements

   • Reshape operation for shared tensors (ReshapeSharedInst, shared_reshape(), inference/validation rules)
   • Canonicalization of SharedLayout (merge consecutive modes, remove singletons, normalize swizzle)
   • Added eq/hash to SharedLayout and Swizzle for structural equality

2. TMA Multicast Support

   • CopyAsyncTensorGlobalToSharedInst and TMA load API now accept optional multicast_mask (uint16)
     to copy data to multiple CTAs in a cluster

3. Cluster API Refactor

   • cluster.block_id() → cluster.blockIdx (property)
   • cluster.shape() → cluster.clusterDim (property)
   • cluster.block_rank() → cluster.blockRank (property)
   • Backed by predefined Var objects instead of inline declare() calls

4. New Hidet Primitives

   • Predefined CUDA variables (clusterBlockIdx, clusterDim, clusterBlockRank, clusterIdx, etc.)
   • BindPredefinedVariablesPass — binds predefined vars to PTX intrinsics at function entry
   • Fence primitives (fence.proxy.async, fence.proxy.tensormap::generic)
   • Integer intrinsics (brev, bfe, bfi, popc, clz, fns, prmt)

5. Misc

   • single_warp() convenience method on RootInstructionGroup
   • flash-attn → flash-attn-4 in CI, updated import paths
   • Blackwell example renamed matmul_v5.py → matmul_v7.py
   • Improved error messages in VectorizedEvaluator
   • Transpiler attribute access simplified (removed hasattr check)
   • Stub flatten() method on RootInstructionGroup (incomplete)