Fix GPT-2 FlashAttention mask preparation without pipeline

[sdjasj]

What changed

• use ColossalAI's GPT-2 model forward whenever FlashAttention is enabled without pipeline parallelism, not only when sequence parallelism is enabled
• let GPT-2 attention-mask preparation handle calls without an explicit mask
• add the reported tp_size=1, pp_size=1, enable_all_optimization=True configuration to the GPT-2 regression matrix

Why

For a non-pipeline, non-sequence-parallel GPT-2 model, the policy replaced GPT2Attention.forward with the ColoAttention implementation but left the Hugging Face GPT2Model.forward unchanged. Hugging Face passed its additive attention-mask Tensor directly to the replacement, which attempted **attention_mask and raised a TypeError because the value was not a mapping.

The existing ColossalAI GPT-2 model forward already converts masks into the keyword dictionary expected by ColoAttention and preserves causal-mask behavior.

Impact

GPT-2 works with HybridParallelPlugin(enable_all_optimization=True) when TP, PP, and SP sizes are all one.

Fixes #6389.

Validation

• 2-GPU minimal GPT-2 training regression with the reported plugin configuration and a padded Tensor mask: forward, backward, and optimizer step passed
• compileall and git diff --check: passed