fix(QwenShardingStrategy): skip weight sharding and ShardedMoE wrap for Qwen3MoeSparseMoeBlock

[FlashShift]

NOTE: This is an AI-produced patch. I have tested it, and it works correctly, but obviously please be cautious with any AI product.

Description:
Problem
Running Qwen3 MoE models (e.g. Qwen3-235B-A22B, Qwen3-397B-A17B) with tensor parallelism across multiple nodes produces a SIGSEGV in libjaccl.dylib inside jaccl::MeshImpl::all_gather, crashing the runner process during prefill.
The crash has two compounding causes, both in QwenShardingStrategy.shard_model:

1. Wrong sharding axis on SwitchLinear weights
   Qwen3MoeSparseMoeBlock.switch_mlp uses SwitchLinear, which stores stacked expert weights as a single 3-D tensor of shape [num_experts, hidden, input]. The all_to_sharded_linear_in_place / sharded_to_all_linear_in_place helpers call shard_inplace, which is designed for nn.Linear and slices on ndim-2 (the output/hidden dimension). For a 3-D stacked weight this is axis 1 (the hidden dimension), not axis 0 (the expert dimension). The result is silently corrupted weight tensors on every rank.
   Qwen3NextSparseMoeBlock and Qwen3_5SparseMoeBlock are not affected — their switch_mlp and shared_expert weights are standard nn.Linear.
2. ShardedMoE wrapping an unsharded block causes a JACCL buffer overrun
   Because the weights are not validly sharded, ShardedMoE.call calls mx.distributed.all_sum on the full unsharded output of each rank. This is mathematically wrong (scales activations by world_size) and — critically — the output tensor during prefill is much larger than the per-token decode tensors JACCL sized its collective buffers for at initialization time. The write overruns the receive buffer, producing:
   SIGSEGV: jaccl::MeshImpl::all_gather(char const*, char*, long long) + 1380
   KERN_INVALID_ADDRESS at 0x0000008210bb0000
   Fix
   Guard both the shard_inplace calls and the ShardedMoE wrap with if not isinstance(layer.mlp, Qwen3MoeSparseMoeBlock). This leaves Qwen3MoeSparseMoeBlock as a plain unsharded module on each rank. Pipeline parallelism handles cross-node communication correctly without any MoE-level collective.
   Testing
   Verified on two Mac Studio nodes (4 GPUs total) running mlx-community/Qwen3-235B-A22B-6bit. Previously crashed immediately on prefill warmup; with this fix inference completes successfully.
   Notes
   A proper tensor-parallel implementation for SwitchLinear-based MoE blocks (sharding on the expert axis) would give memory reduction benefits and is worth a follow-up, but requires a custom collective that understands the 3-D weight layout and the sparse routing indices.