optimization='auto' cannot select metal or cuda_kernel unless torch is installed; availability probes never check for a device

[Ramdam17]

1. Platform detection depends entirely on torch.

MPS_AVAILABLE and CUDA_AVAILABLE are derived only from torch (hypyp/sync/base.py:53-61):

try:
    import torch
    TORCH_AVAILABLE = True
    MPS_AVAILABLE = torch.backends.mps.is_available()
    CUDA_AVAILABLE = torch.cuda.is_available()
except ImportError:
    TORCH_AVAILABLE = False
    MPS_AVAILABLE = False
    CUDA_AVAILABLE = False

_resolve_auto picks its platform from those two flags (base.py:414-418). So on a machine with CuPy and an NVIDIA GPU but no torch, optimization='auto' warns "No GPU available" and falls back to numba — even though cuda_kernel is present and working. Same for a Mac with pyobjc-framework-Metal but no torch: 'auto' can never reach the Metal kernels.

The two custom-kernel backends are therefore unreachable via 'auto' without installing torch, which is not a dependency of either. pip install "hypyp[cupy]" alone gets you a CPU fallback.

2. Availability probes are import-only and never touch the device.

hypyp/sync/kernels/__init__.py:10-21 sets both flags purely on import success:

• METAL_AVAILABLE = True iff import Metal succeeds — MTLCreateSystemDefaultDevice() is never called. If it returns None (headless/VM/unsupported GPU), the first compute() dies on AttributeError: 'NoneType' object has no attribute 'newLibraryWithSource_options_error_'.
• CUPY_AVAILABLE = True iff import cupy succeeds — cupy.cuda.runtime.getDeviceCount() is never called. On a cluster login node with CuPy installed but no GPU allocated (e.g. Narval), optimization='cuda_kernel' promises a GPU it does not have, and the 9 CUDA tests run and error at kernel launch instead of skipping.

Note the asymmetry: CUDA_AVAILABLE does a real runtime check via torch.cuda.is_available(), but CUPY_AVAILABLE does not.

Suggested fix: derive platform availability independently of torch (probe Metal device creation and cupy.cuda.runtime.getDeviceCount() directly), and make both flags reflect an actually usable device rather than a successful import.