[flex_shard] Add DeepSeek V3 eager training entry point

[weifengpy]

Stack from ghstack (oldest at bottom):

• Add communication-free Muon for FlexShard #3502
• [flex_shard] Plan: fp8 all-gather on GroupedRaggedShard (block-wise) #3537
• -> [flex_shard] Add DeepSeek V3 eager training entry point #3603
• Make FlexShard traceable by torch.compile #3317
• Introduce FlexShard for flexible bucketed parameter sharding #3239

Add an experimental flex_shard.deepseek_v3 module that reuses the DeepSeek V3 model config with a FlexShard eager parallelizer. The path slices local experts for EP, shards dense params on the FSDP mesh, shards local expert params on the expert-FSDP mesh, and rejects unsupported PP/TP/CP/HSDP/model-compile modes for now.

Make FlexShard work with TorchTitan's meta-module to_empty flow by deferring eager hook installation until bucket storage is materialized, then reinstalling sharded parameter views. Also add FlexShard handling for EP grad norm clipping and DCP state-dict introspection.

Include runtime/state-dict coverage plus local design notes and repro/profiling helpers for the ongoing FlexShard training and compile investigations.

Test Plan:

• CUDA_VISIBLE_DEVICES=6 NGPU=1 MODULE=deepseek_v3 CONFIG=deepseek_v3_debugmodel_ep ./run_train.sh --parallelism.data_parallel_shard_degree=1 --parallelism.expert_parallel_degree=1 --training.steps=1 --activation_checkpoint.mode=none --dump_folder=outputs/retry_fully_shard_norm_dp1

• CUDA_VISIBLE_DEVICES=0,1,6,7 NGPU=4 MODULE=deepseek_v3 CONFIG=deepseek_v3_debugmodel_ep ./run_train.sh --parallelism.data_parallel_shard_degree=4 --parallelism.expert_parallel_degree=4 --training.steps=1 --activation_checkpoint.mode=none --dump_folder=outputs/retry_fully_shard_norm_dp4_ep4