feat: CUDA 10.2 / C++14 compatibility for Jetson TX2 (compute 6.2)

[sourceupdev]

THIS IS A CHANGE ONLY TO PERSIST A WORKING VERSION ON JETSON-TX2, not intended for MAIN Llama-CPP

Minimal-diff approach to support GCC 9 + nvcc 10.2 with --expt-relaxed-constexpr:

• Add compat-cuda10.cuh: bf16->fp16 polyfills (no hw bf16 on compute 6.2)
• Guard cuda_bf16.h include with CUDART_VERSION >= 11000
• CMake: C++14 std, arch 62, --expt-relaxed-constexpr for CUDA < 11.0
• Replace std::is_same_v with std::is_same<>::value (C++14)
• Convert fold expressions to C++14 equivalents
• Convert structured bindings to explicit .first/.second
• Guard cooperative_groups (cg::this_grid) behind CUDART_VERSION >= 11000
• Fix cudaStreamWaitEvent 2-arg calls (CUDA 10.2 requires 3rd flags param)
• Replace __builtin_assume with GGML_CUDA_ASSUME macro
• Fix static inline const/auto to constexpr with explicit types
• Fix if-init statements to C++14 style

14 files changed (13 modified + 1 new), 160 insertions, 98 deletions Tested: build OK, CPU inference 0.7 t/s, GPU inference 7.3 t/s (gemma-4-E2B Q4_0)

Overview

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: