add metal-flash-attention install script and docs

[bghira]

This pull request updates the experimental Universal Metal Flash Attention (UMFA) documentation in multiple languages to clarify requirements for macOS and toolchains, and to add instructions for using new helper scripts in SimpleTuner. The changes improve clarity, add automation for setup, and ensure consistency across English, Spanish, Hindi, Japanese, Portuguese, and Chinese documentation.

Key documentation improvements:

Requirements and prerequisites:

• Updated minimum requirements to specify macOS 15+, Apple Silicon, Git, Swift 6+, and Xcode 16+ (or equivalent command line tools with the Metal toolchain) in all language variants. [1] [2] [3] [4] [5] [6]

Helper script usage and automation:

• Added instructions for using the new apple-metal-flash-attention.sh helper script in SimpleTuner to check the local environment and automate the clone/build/install process for UMFA. This includes both --check and --install usage examples. [1] [2] [3] [4] [5] [6]
• Described the steps performed by the install path: cloning UMFA if needed, initializing submodules, building the Swift product MFAFFI, building the PyTorch custom-op extension in place, and installing the Python binding. [1] [2] [3] [4] [5] [6]

Manual build instructions:

• Updated Swift build instructions to explicitly build the MFAFFI product. [1] [2] [3] [4] [5] [6]
• Added explicit step to build the PyTorch extension in place with setup.py build_ext --inplace. [1] [2] [3] [4] [5] [6]

These changes ensure that users in all supported languages have clear, up-to-date, and consistent instructions for setting up UMFA with SimpleTuner.