[attention] Make the FlexAttention mask type configurable

[aws-arashj]

Summary

FlexAttention always builds a block-causal mask, which is causal within each packed document. This PR adds FlexInnerAttention.Config.mask_type so a model can choose plain "causal" instead, matching the transformers backend's attn_mask_type.

• "block_causal" is the default and keeps today's behavior.
• "causal" doesn't depend on positions, so the mask is the same for every batch. It's only correct when each sequence holds one document.

A new helper, get_flex_causal_mask_mods(positions, config), returns the selected mask_mods. The decoder, GPT-OSS, and Muse Glimmer mask builders use it instead of composing the masks inline. Masks are unchanged for every existing config.

Test plan

• New tests/unit_tests/cpu/test_flex_mask_type.py (2 tests, 9 subtests, all passing):
  • mask_mods. On packed and single-document positions, block_causal is causal ∧ same-document and causal is causal. The two agree on single-document sequences.
  • Mask builders. For the decoder, GPT-OSS, and Muse Glimmer, every mask returned with "causal" equals the lower-triangular block mask, and the default block_causal keeps fewer blocks. That covers both the full and sliding-window masks. It fails if either model's builder ignores mask_type.
• test_flex_attention.py, test_gpt_oss_moe.py, test_muse_glimmer.py, and test_varlen_attention.py: 13 passed, the same as on main.