Bug Report: RTX 5090 incompatible with default Pinned Memory

## Bug Report: RTX 5090 incompatible with default Pinned Memory

**Environment:**
- GPU: NVIDIA GeForce RTX 5090 Laptop (24GB)
- OS: Windows
- CUDA: 12.4+
- PyTorch: 2.x

**Root Cause:**
The new default-enabled Pinned Memory calls `cudaHostRegister(ptr, size, 1)`
which returns "CUDA error: invalid argument" on Blackwell architecture GPUs.

**Fix Suggestion:**
1. Change `cudaHostRegister` flag from `1` to `0` (or use `cudaHostRegisterDefault`)
2. Add error handling to disable Pinned Memory if registration fails
3. Add compute capability check for Blackwell (>= 10.0)

**Workaround:**
Add `--disable-pinned-memory` to launch arguments.

Locate the pinned-memory implementation and the cudaHostRegister(ptr, size, 1) call; no file or test is named in the report. Reproduce on the listed RTX 5090 Windows environment, then evaluate the suggested flag change, registration-failure handling, and Blackwell compute-capability check. Done means pinned memory no longer causes the reported invalid-argument error, while the documented workaround remains available.

Written by the indexing model from the issue text.