Add CPU and CUDA autocast scopes backed by LibTorch

[lunosia]

Summary

• Expose torch.autocast(DeviceType, ScalarType) and is_autocast_enabled for CPU and CUDA through LibTorch.
• Preserve thread-local enabled state and dtype across nested scopes. Clear cached casts when the dtype changes so BF16 and FP16 scopes can nest safely.
• Add CPU and CUDA tests for output dtype, FP32 parameters and gradients, optimizer updates, nesting, and thread-affine disposal. The Linux CUDA test project resolves its runtime package directly.
• On Windows with LibTorch 2.10, remove whole-archive only from the torch.lib and torch_cpu.lib DLL import libraries. Without this, the native build fails with duplicate import descriptors (LNK2005/LNK1169).

Validation

• Windows x64 / Visual Studio 2022: native and managed build succeeded; net8.0 CPU suite 1332 passed, 30 skipped; net472 test project compiled.
• Ubuntu x64 / RTX 3090 / CUDA 12.8 runtime: CPU suite 1329 passed, 33 skipped; CUDA runtime suite 1324 passed, 26 skipped.
• With CUDA hidden, the two CUDA autocast tests fail explicitly instead of reporting a false pass.
• A downstream C# language-model project completed BF16 autocast short training and checkpoint resume on the RTX 3090.

Scope and related work

• This exposes BF16 and FP16 autocast scopes. It does not add a GradScaler for FP16 training.
• Related to Autocast #1235, which explores broader AMP support. This PR keeps the API and tests focused on the existing LibTorch autocast implementation.

[lunosia]

@dotnet-policy-service agree