[PyTorch] Add record_stream and untyped_storage func op in QuantizedTensor

[xiaoxi-wangfj]

Description

In the FP8 dataflow (dispatch → expert fc1) path, under 1F1B overlap, we need to switch to the work stream and safely release memory from the previous stream when passing a QuantizedTensor (FP8 payload plus scale_inv metadata) from dispatch to expert fc1. This reduces peak HBM usage and avoids unnecessary memory retention.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• torch_dispatch handler for aten.record_stream on QuantizedTensor:
  We record all relevant CUDA buffers inside the quantized tensor—_rowwise_data/_columnwise_data and their _rowwise_scale_inv/_columnwise_scale_inv—onto the provided stream via record_stream(stream). This does not change tensor values; it only updates storage lifetime metadata so the allocator won’t reuse/free the memory before the stream finishes its asynchronous work.

• Expose QuantizedTensor.untyped_storage():
  Returns the payload’s underlying UntypedStorage. Callers can then run resize_(0) to immediately shrink the storage capacity to zero and return it to the caching allocator (on CUDA).

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes