Add Safetensors.write!/2 for streamed write

[jonatanklosko]

When persisting to a file, Safetensors.dump/1 calls Nx.to_binary/1 on all tensors upfront (and if we need to convert endianness, we create yet another binary), so this can cause a significant memory bump. This PR adds Safetensors.write!/2 that converts and writes tensors directly to the file, one by one.

Other than lower memory usage, this can also be much faster. For the example below, it's 5 times faster for me:

tensors =
  # 1000 tensors, 8MB each -> 8GB
  for i <- 1..1000, into: %{} do
    {"t#{i}", Nx.iota({1_000_000}, type: :f64)}
  end

# Dump + write
File.write!("/tmp/test.safetensors", Safetensors.dump(tensors))

# Direct write
Safetensors.write!("/tmp/test.safetensors", tensors)

[devCaiquedePaula]

In lib/safetensors. The Elixir module lacks comprehensive documentation, particularly at the module level, where an explanation of the module's purpose and key concepts would be beneficial. Consider improving code consistency by ensuring uniform indentation and adding vertical spacing for better readability. Address error handling, especially in the File.open!/2 call, by considering alternative approaches such as using File.open/2 to handle potential errors. Refactor code duplication between the write!/2 and dump/1 functions to a common function to enhance maintainability. Introduce named constants or functions for magic numbers and repeated calculations. Validate input parameters, improve error messages for better user guidance, and handle file closing after usage. Verify type specifications for accuracy, and ensure thorough testing to validate the correctness of implemented functions. Overall, these refinements will enhance code organization, maintainability, and user experience.

[josevalim]

Hi @devCaiquedePaula. Please do not use automated/AI code review tools unless requested. If you feel any of these points are relevant though, a PR is welcome.