ggml : 4-bit Integer quantisation + many llama.cpp improvements (#27)

ggml : 4-bit Integer quantisation + many llama.cpp improvements (#27)

* gq : attempt at n-bit quantization

* gq : add amax based method 3

* gq : progress on method 2

* gq : method 4 (AVX2)

* gq : method 4 (ARM)

* gq : method 4 (AVX2 attempt) + method 5 (no min)

* gq : method 5 (ARM)

* gpt-2 : model conversion for Q4_0 quantization

* ggml : Q4_0 quantization support (ggml_get_rows())

* gpt-2 : loading Q4_0 quantized model

* ggml : q4_0 quantization support

* ggml : q4_1 quantization support (seems to work for bigger models)

* gpt-2 : add gpt-2-quantize tool for quantizing f32 GPT-2 models

* ggml : 4-bit quantization works (only scalar for now)

* gq : add method 6 (ARM)

* ggml : vectorized mad q4_0 (ARM)

* ggml : vectorized quantize_row_q4_0 (ARM)

* ggml : simplify mad q4_0 (ARM)

* ggml : minor indentations

* gpt-j : support for 4-bit quantized model inference

* ggml : GGML_ASSERT() instead of assert() where appropriate

* gpt : avoid ggml_transpose on model tensors (new models!)

* gpt-2 : minor

* gpt-j : fix conversion for FP16 models (such as GPT-JT-6B)

* ggml : add ggml_compute_forward_rope_f16()

* gpt : fix memory usage computation

* ggml : fix ggml_is_contiguous() to take into account blck size

* whisper : add whisper-qunatize tool

* whisper : add support for quantized models

* whisper : mem usage based on model format type

* gpt : seems not worth to use FP16 for KV cache

* gpt : support quantisation of f16 models files

* ggml : fixes for rpi4

* whisper : add Q4_1 model sizes

* ggml : add WASM SIMD for Q4_0

* utils : print quantization histograms

* ggml : sync all changes from llama.cpp and whisper.cpp

* ggml : finalize the Q4_1 quantization for ARM_NEON