Apply lintrunner, and stop clang-format breaking the FlashMLA launch
`lintrunner -a` runs at the top of the build script, and on `Name<<<...>>>` with a
non-template kernel clang-format rewrites the closing brackets to `> > >`, which
does not compile. The broken form was already committed, so the tree did not build
once ORT_ENABLE_FLASH_MLA started pulling flash_mla_latent.cu in. Guard that one
launch and take the formatter's output for the rest.