ggml : make ggml_fp16_t private

[ggerganov]

Currently the ggml_fp16_t typedef is exposed in the public API:

ggml/include/ggml/ggml.h

Lines 317 to 332 in 6b14d73

	#if defined(__ARM_NEON) && defined(__CUDACC__)
	typedef half ggml_fp16_t;
	#elif defined(__ARM_NEON) && !defined(_MSC_VER)
	typedef __fp16 ggml_fp16_t;
	#else
	typedef uint16_t ggml_fp16_t;
	#endif
	// convert FP16 <-> FP32
	GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x);
	GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
	GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);
	GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);

Since this type is platform specific, it would make sense to hide it by moving it in ggml-impl.h.
We will still expose an API for F16 <-> F32 conversions, but it sill operate on void * instead of ggml_fp16_t

[slaren]

We could probably define it to uint16_t always, and only cast it to __fp16 in the ARM code that can take advantage of that. The CUDA one is weird, it should only be used when compiling ggml-cuda.cu, which never uses this type anyway.

[ggerganov]

Ah yes, doing so would make this change much simpler