feat: Add download acceleration for dependencies and HuggingFace models

[deanq]

Introduces support for download acceleration from Tetra runtime. This also speeds up remote execution startup times by pre-caching pip dependencies and HuggingFace models.

• Parallel downloads for large files
• Applies accelerated downloads to known large pip libraries and hugging face models
• Smart-caching of hugging face models in the container or network volume

Includes new accelerate_downloads and hf_models_to_cache parameters for the @remote decorator with backward compatibility.

# Accelerated downloads are on by default. To turn it off...
from tetra_rp import remote, LiveServerless

gpu_config = LiveServerless(name="my_server")

@remote(
    resource_config=gpu_config,
    dependencies=["vllm"],
    accelerate_downloads=False,
)
...

# cache the model to the container after the first time it is downloaded
from tetra_rp import remote, LiveServerless

gpu_config = LiveServerless(name="my_server")

@remote(
    resource_config=gpu_config,
    dependencies=["vllm"],
    hf_models_to_cache=["facebook/opt-125m"],
)
...

# cache the model to the volume after the first time it is downloaded
from tetra_rp import remote, LiveServerless, NetworkVolume

gpu_config = LiveServerless(
    name="my_server",
    networkVolume=NetworkVolume(name="my_volume")
)

@remote(
    resource_config=gpu_config,
    dependencies=["diffusers", "torch", "transformers", "accelerate", "xformers"],
    hf_models_to_cache=["runwayml/stable-diffusion-v1-5"]
)
...

Related to runpod-workers/worker-tetra#22

[pandyamarut]

/LGTM