fix: forward cache_dir to ONNX and OpenVINO model loading

[AtharvaJaiswal005]

Summary

• Forward cache_dir from Transformer._load_model() to load_onnx_model() and load_openvino_model()
• Forward cache_dir to the underlying model_cls.from_pretrained() calls in both functions
• Fixes ONNX/OpenVINO backend failing with ValueError when cache_folder is specified alongside local_files_only=True

Root Cause

When using backend="onnx" (or "openvino") with a custom cache_folder, the cache_dir parameter was not forwarded through the loading chain:

1. Transformer._load_model() receives cache_dir but did not pass it to load_onnx_model() / load_openvino_model() (unlike the torch backend which passes cache_dir to AutoModel.from_pretrained())
2. load_onnx_model() / load_openvino_model() did not accept or forward cache_dir to model_cls.from_pretrained()

This meant the ONNX/OpenVINO model loader had no knowledge of where cached files were stored, causing it to fail when local_files_only=True.

Reproduction

from sentence_transformers import SentenceTransformer

# Step 1: Download to cache (works)
model = SentenceTransformer(
    "sentence-transformers/all-MiniLM-L6-v2",
    cache_folder="./model_cache/",
    backend="onnx",
    local_files_only=False,
)

# Step 2: Load from cache (fails before fix)
model = SentenceTransformer(
    "sentence-transformers/all-MiniLM-L6-v2",
    cache_folder="./model_cache/",
    backend="onnx",
    local_files_only=True,
)

Changes

• sentence_transformers/backend/load.py: Added cache_dir parameter to load_onnx_model() and load_openvino_model(), forwarded to from_pretrained()
• sentence_transformers/models/Transformer.py: Pass cache_dir when calling load_onnx_model() and load_openvino_model() in _load_model()

Fixes #3562