Fix DeepSeek related bugs (#724, #548, #633)

[Nyquist24]

• Fix pretrain_deepseek.py 136行有一个没有定义的函数，搜索仓库根本没有 get_gpt_decoder_layer_specs #724: Remove undefined function get_gpt_decoder_layer_specs call in pretrain_deepseek.py, reuse already-computed transformer_layer_spec
• Fix example deepseek_v2 flash_attn option is useless #548: Remove non-functional flash_attn option in deepseek_v2 since MLA does not support flash attention, always use TE Fused Attention
• Fix hf2mcore转换后，模型缺少local_experts #633: Replace unsafe str.replace with re.sub(count=1) in hf2mcore_deepseek_v2_moe.py to prevent substring mismatch during local_experts key renaming

Changes

File	Change
examples/deepseek_v3/pretrain_deepseek.py	Remove 10 lines of redundant/broken MTP spec code
examples/deepseek_v2/run_mcore_deepseek.sh	Replace FL conditional with fixed NVTE_FUSED_ATTN=1
toolkits/.../hf2mcore_deepseek_v2_moe.py	3x str.replace → re.sub(..., count=1)

Test plan

• Verify DeepSeek-V3 pretraining with --mtp-num-layers runs without NameError
• Verify DeepSeek-V2 training uses fused attention correctly
• Verify HF→mcore MoE checkpoint conversion with EP>1 produces correct local_experts keys

[CLAassistant]

All committers have signed the CLA.