Universal CUDA/ROCm wheel via runtime platform detection

[wjabbour]

Summary

• Replaces compile-time ROCM_PATH / USE_ROCM platform selection with runtime dlopen detection
• On startup, tries libcudart.so first, then libamdhip64.so — whichever succeeds sets the GPU function pointers
• is_cuda_found() and is_hip_found() are now real runtime queries (both always exported), not compile-time constants
• is_gds_supported() uses the runtime is_hip_runtime flag instead of #ifdef USE_ROCM
• Renames cuda_compat.h → gpu_compat.h with updated include guards and two parallel CUDA_SYM_* / HIP_SYM_* define sets (no #ifdef guards)
• Removes detect_platform() from setup.py — single build works on both platforms
• Adds missing is_hip_found() stub to cpp.pyi
• Updates README ROCm install instructions (no longer need ROCM_PATH= prefix)

Test plan

• Built and installed locally on ROCm machine (RDNA 4)
• is_cuda_found() → False, is_hip_found() → True on ROCm
• Loaded Qwen3-0.6B (311 tensors, single shard) to cuda:0
• Loaded Hermes-3-Llama-3.1-8B (291 tensors, 4 shards) to cuda:0 in 9.5s
• vLLM --load-format fastsafetensors with ibm-granite/granite-3.3-2b-instruct on ROCm (weights loaded in 0.43s, server fully up)
• CUDA machine smoke test (CI)

Closes #68

🤖 Generated with Claude Code

[wjabbour]

@takeshi-yoshimura Hey Takeshi, this PR is good for review!

[takeshi-yoshimura]

@wjabbour
Cool! This is exactly what I envisioned while ago to support both ROCm and CUDA within a single package. Thanks.