Fix memory allocation for base-patch16

[monatis]

Closes #17

[Green-Sky]

confirm working

clip_model_load: loading model from '../models/openai_clip-vit-base-patch16/openai_clip-vit-base-patch16.ggmlv0.f16.bin' - please wait...
clip_model_load: text model hparams
n_vocab            49408
num_positions      77
t_hidden_size      512
t_n_intermediate   2048
t_projection_dim   512
t_n_head           8
t_n_layer          12

clip_model_load: vision model hparams
image_size         224
patch_size         16
v_hidden_size      768
v_n_intermediate   3072
v_projection_dim   512
v_n_head           12
v_n_layer          12

use_gelu           0
ftype              1

clip_model_load: ggml ctx size = 287.12 MB
.................................................clip_model_load: model size =   285.77 MB / num tensors = 397
clip_model_load: 16 MB of compute buffer allocated
clip_model_load: model loadded

apple = 0.1522
red = 0.1502
human = 0.1427
dog = 0.1404
photo = 0.1388
blue = 0.1381
drawing = 0.1376

[Green-Sky]

why use == here and <= down below.
also why different param name?

[monatis]

We call this function only once in clip_model_load and allocate that memory based the number of positions in the vision model. The scratch buffer, however, is allocated separately in clip_image_encode and clip_text_encode. n_image_positions is fixed for a gien model, but n_positions might be different in the case of the text model. So if I used == there, the condition would not return true for shorter texts.