Fix ROCm compatibility for HIPBLAS and Thrust includes

[vinayakv22]

This PR fixes ROCm build issues observed when compiling on gfx1031 (RX 6800M).

Changes

• Fix HIPBLAS symbol mapping by removing unsupported _v2 aliases:
  • cublasGemmEx -> hipblasGemmEx
  • cublasGemmStridedBatchedEx -> hipblasGemmStridedBatchedEx
• Add missing Thrust headers used by CUDA/HIP codepaths:
  • thrust/reduce.h, thrust/extrema.h
  • thrust/iterator/counting_iterator.h
  • iterator headers in helpers.h and gumbel_max_gpu.cu

Result

• libctranslate2.so now builds successfully on ROCm in this environment.
• Build still reports an unrelated CLI compile issue (cxxopts missing uint8_t include), but core library target is successful.