Add pwrite-based single-file ONNX export + fix GPT-OSS dynamo export

[smedhe]

Summary

• Replaces onnx_program.save() with a new save_dynamo_onnx() helper (QEfficient/utils/onnx_save_utils.py) that writes all initializer tensors in parallel via os.pwrite into a single pre-allocated .data file, then serializes a small graph-only .onnx
  proto referencing them via ExternalTensor offsets. This avoids the ~17M small write() syscalls and per-file inode overhead of the default save path, which is especially costly on large models on HDD-backed storage. Includes disk-space precheck, atomic
  rename via .data.tmp, per-tensor byte-count verification, RAM-headroom warning (advisory, psutil optional), and an opt-in sync flag for fdatasync durability.
• Wires the new save path into QEFFBaseModel's dynamo export flow (modeling_qeff.py), and explicitly del onnx_program; gc.collect() afterward to release ExportedProgram weakrefs so PT weight offload can proceed.
• Fixes GPT-OSS dynamo export/compile issues in modeling_gpt_oss.py (rotate_half rewritten using torch.chunk) and cache_utils.py (QEffHybridCacheForGPTOSS — drops redundant torch._dynamo.is_compiling() int-cast branching for
  max_cache_len/sliding_window_len).
• Adds a dynamo: bool = False parameter to QEFFAutoModelForCausalLM.compile()/export plumbing (modeling_auto.py) to opt into the dynamo-based export path, and stores self.onnx_path on the base model for later reference.

Why

The default onnx_program.save() serializes weights through a BufferedWriter with an 8KB buffer, resulting in millions of small writes — this becomes a major bottleneck for large models (e.g. 70B+) on spinning-disk-backed storage. The single-file pwrite
approach lets the OS page cache coalesce writes into large sequential flushes, cutting save time from ~45 min to ~2 min for a 131GB model in local testing. The GPT-OSS fixes address failures encountered when exporting that model through the dynamo path.