fix: skip pin_memory and non_blocking transfer for tensor subclasses in group offloading

[s-zx]

What does this PR do?

Fixes #13281.

When combining torchao quantization (e.g. Float8WeightOnlyConfig) with group offloading (use_stream=True), inference fails with a device mismatch: the quantized weight remains on CPU while the input is on CUDA.

Root cause

GroupOffloadingHook._init_cpu_param_dict calls .pin_memory() on every parameter/buffer, and _pinned_memory_tensors pins them again before the async host-to-device transfer. For plain torch.Tensor this works correctly, but for tensor subclasses such as torchao's AffineQuantizedTensor, pin_memory() may silently strip the quantization metadata, producing a plain CPU float tensor. The subsequent .to(device, non_blocking=True) then moves a plain float tensor instead of the quantized one, leaving the quantized weight behind on CPU.

Similarly, _transfer_tensor_to_device schedules record_stream for all results, but AffineQuantizedTensor.to(...) may not return a plain torch.Tensor, so record_stream could raise or be a no-op.

Fix

• _init_cpu_param_dict: introduce _maybe_pin() helper that calls pin_memory() only when type(data) is torch.Tensor. Tensor subclasses are stored as-is (already on CPU).
• _pinned_memory_tensors: apply the same type(tensor) is torch.Tensor guard.
• _transfer_tensor_to_device: force synchronous (blocking) transfer for tensor subclasses, and skip record_stream when the result is not a plain tensor.

Plain-tensor performance is completely unaffected; the async-stream path is preserved for all non-quantized parameters.