Introduce set_default_sdpa

[cbalioglu]

This PR introduces set_default_sdpa function and sdpa context manager to switch between different attention implementations during runtime.

from fairseq2.nn.transformer import TorchSDPA, NaiveSDPA, set_default_sdpa, sdpa

set_default_sdpa(TorchSDPA)  # or None to use the library default

# Use naive SDPA for debugging (e.g. pdb)
with sdpa(NaiveSDPA)
    model = load_llama_model("llama_7b")