Make CUDA runtime init soft-fail when no device is present

[badnikhil]

Follows up on #100 - ensure the runtime module does not cause issues trying to load itself when no CUDA is present.

The guard from #100 doesn't achieve this: it wraps runtime.d's module constructors in version(LDC_DCompute_CUDA), but no compiler defines that identifier — LDC predefines only LDC_DCompute The constructors are therefore compiled out in every build, and on machines that do have a GPU the default context is never created:

Changes

• New Platform.tryInitialise(): a soft-failing probe that never throws and never mutates the thread-local error status - checks loadCUDA() (noLibrary/badLibrary), the cuInit result, and cuDeviceGetCount > 0.
• _initPlatform() marks CUDA unavailable and returns quietly when the probe fails; device/context creation is wrapped in a defensive try/catch so nothing can escape shared static this() and abort the process before main().
• Per-thread static this() no-ops when the platform isn't ready.
• ensureInit() now throws a catchable DComputeDriverException("No CUDA driver or device available on this system.") when CUDA is absent, so launch!'s lazy-init path reports cleanly instead of failing later with a confusing invalidContext.

Verification (RTX 2050; device masking via CUDA_VISIBLE_DEVICES)

Scenario	Before	After
CUDA build, GPU present	suite fails (invalidContext, exit 1)	unmodified suite passes (exit 0)
CUDA build, device masked ("" / -1)	constructors dead, invalidContext on first Buffer	survives constructors; clear catchable error on first CUDA use
OpenCL build (also compiles cuda/runtime.d), CUDA masked	—	constructors run and soft-fail quietly; program proceeds into OpenCL host code

An opt-in automated test in tests/runtime_init.d (env-guarded) covers the no-device path: DCOMPUTE_TEST_NO_CUDA=1 CUDA_VISIBLE_DEVICES="" ./dcompute verifies the constructors survive and ensureInit() throws as expected.