[FIX][ROCm] Use HIP for device existence checks

[oraluben]

ROCm device existence checks directly include hsa/hsa.h and call HSA initialization/shutdown around a HIP query. The tested Windows ROCm 10.0 pip SDK omits that header, preventing compilation, although its HIP DLL exports working HSA functions.

Use hipGetDeviceCount on Linux and Windows. HIP automatically initializes its runtime; on the ROCr path this includes HSA initialization and cleanup. Removing TVM's extra HSA reference does not bypass that initialization. Return false on HIP errors or invalid ordinals, remove unused HSA discovery/linkage, and add portable device-query regressions.

Validation:

• Windows: actual upstream runtime source compiles with the selected SDK. The downstream mirror passes all 5 device tests and 10 GPU numerical tests on both TileLang adapters. A standalone process, with no Torch or explicit HIP/HSA initialization, passes the first device query, 400 concurrent queries, allocation and GPU copies.
• Linux/WSL: built both upstream ROCm runtime targets and the mirror runtime against TileLang's HIP stub; actual device API queries safely report absent/invalid devices. The consumer's 12 stub regressions and real ELF symbol-reuse probe pass. Also checked the implicit initialization path in ROCm 6.2.4 sources.

Linux GPU execution and a full upstream compiler build were not tested. Mirror: tile-ai#79. Consumer: tile-ai/tilelang#3425.