Documentation: Accelerate executor with Flash Attention

[l45k]

To use Flash Attention in the executor, it needs to be started with Flash Attention as an additional dependency like:

[[executors]]
class = "train"
name = "diloco-transformer"
runtime = "process"
cmd = "uv"
args = [
    "run",
    "--python",
    "3.12",
    "--no-project",
    "--with", "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.2/flash_attn-2.8.3+cu128torch2.8-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl",
    "--",
    "accelerate",
    "launch",
    "--config-file",
    "accelerate.yaml",
    "-m", "hypha.accelerate_executor.training",
    "--socket",
    "{SOCKET_PATH}",
    "--work-dir",
    "{WORK_DIR}",
    "--job",
    "{JOB_JSON}",
]