CMSIS-DSP 1.18.0

• Includes the optional C++ template headers in the CMSIS-DSP pack. They provide a higher-level API with automatic loop fusion for improved performance.
• Added experimental reverse-mode automatic differentiation with f32 and f16 support for on-device fine-tuning using only existing CMSIS-DSP kernels. This is not a new machine-learning framework; the focus is learning parameters of selected CMSIS-DSP kernels.
• Added examples for the C++ API
• Corrected Q15 FIR state buffer sizing when ARM_MATH_LOOPUNROLL is enabled (#305, #317).
• Removed implementation-defined signed shifts from fixed-point clipping helpers (#211) and expanded their tests.
• Corrected the INT64_MIN / -1 overflow case in arm_div_int64_to_int32 (#322).
• Corrected normalization by 32 bits in arm_norm_64_to_32u.
• Corrected controller benchmark tests (#325).
• Documented the non-saturating final conversion in arm_pid_q31 (#327).