[FlyDSL]: MI355 FMHA kernel refactor for modularity and code simplification

[yanguahe]

Suggestion Description

Refactor the MI355X FMHA kernel to make the implementation more modular, remove redundant code, and reduce the overall line count.

The refactor should use official FlyDSL APIs wherever possible, align the code with FlyDSL conventions, and make the kernel easier to maintain, extend, and tune.

Scope

• Split reusable MI355X FMHA logic into clearer modules.
• Remove duplicated logic and reduce the total code size.
• Prefer official FlyDSL APIs over non-standard or temporary implementations.
• Preserve the existing functionality and performance baseline while avoiding behavioral regressions.

[yanguahe]

Current progress: the corresponding PR has been opened: #845 (#845).

The refactor work is in progress on branch hyg/fmha_refactor and currently includes 7 commits. The main changes extract shared MI355/GFX950 FMHA helpers into dualwave_common, parameterize the dualwave SWP traits, move load/mask/MMA/rescale/store helper logic out of flash_attn_gfx950, and slim the GFX950 kernel file down toward orchestration and entry wiring.

Current status: the PR is open, CodeQL checks are passing, and review is still required before merge.