Implement the YaRN rop scaling feature

[xiguiw]

Add new API for YaRN rop-scale:

NE_API struct ne_tensor* ne_rope_custom_inplace(struct ne_context* ctx, struct ne_tensor* a, int n_past, int n_dims, int mode,
int prompt_size, float freq_base, float freq_scale,
int yarn_orig_ctx, float ext_factor, float attn_factor,
float beta_fast, float beta_slow);

// shift all tokens by a give p (n_shift)
// Optionally give a 1d tensor of precomputed interleaved cos/sin value of n_shiftscale^k for k \in [0, n_dims)
NE_API struct ne_tensor ne_rope_custom_shift_inplace(struct ne_context* ctx, struct ne_tensor* a, int n_shift, int n_dims,
int mode, int prompt_size, int n_keep, struct ne_tensor* cossin,
float freq_base, float freq_scale,
int yarn_orig_ctx, float ext_factor, float attn_factor,
float beta_fast, float beta_slow);

Change ne_layer internal API:
original:
struct ne_tensor* ne_rope_impl(struct ne_context* ctx, struct ne_tensor* a, int n_past, int n_dims, int mode,
int prompt_size, bool inplace, int n_keep, struct ne_tensor* cossin, int* n_padding,
bool padding_left, float freq_base, float freq_scale)
new API:
struct ne_tensor* ne_rope_impl(struct ne_context* ctx, struct ne_tensor* a, int n_past, int n_dims, int mode,
int prompt_size, bool inplace, int n_keep, struct ne_tensor* cossin, int* n_padding,
bool padding_left, float freq_base, float freq_scale,
int yarn_orig_ctx, float ext_factor, float attn_factor,
float beta_fast, float beta_slow)

model calling ne_rope_impl, it's behavior should not be changed.

[airMeng]

you can include related python scripts update in this pr

neural-speed/neural_speed/convert/convert_baichuan.py

Line 163 in 5a90dc4

fout.write(struct.pack("f", 0.0)) # config.json "rope_scaling.factor", not enabled

[intellinjun]

https://inteltf-jenk.sh.intel.com/job/neural_speed_extension/49/
This ci test llama2 and gptneox, check if this PR affects the original model that uses rope.

[airMeng]

fix the format issues then merge

[intellinjun]

https://inteltf-jenk.sh.intel.com/job/neural_speed_extension/50/ This ci test llama2 and gptneox, check if this PR affects the original model that uses rope.

please wait result before merge.

[DDEle]

No difference. The only difference is that b->data can set size case by case while op params have a shard max size as NE_MAX_OP_PARAMS. In addition, ->data was used as a workaround before op_param came.

Remove / modify this comment if no further questions on this.