Fix embeddings extraction

[skirodev]

This is a pull request to fix some errors in the "extract embeddings" section of the original code in the repository. The running results of the code after the corrections were made when obtaining embeddings are shown below.

The Result of calculating the cosine similarity using the fixed code and llama.cpp:
Query Text: ”My favourite animal is the dog“

Text	Similarity(LLaMA.cpp)	Similarity(Fixed Code)
“My favourite animal is the dog”	1	0.99999999
“My favourite animal is the cat”	0.928901	0.92852515
”I have just adopted a cute dog“	0.70601	0.596631

The first 10 elements of text embeddings, the length of all embeddings is 4096:

1. Query Text: "My favourite animal is the dog"
   Fixed Code:
   [1.6758084, -0.10050424, -0.3598353, -4.8037696, 1.4627854, -2.3149424, 2.0896077, -1.4845643, 0.5678074, -0.18290481]
   LLaMA.cpp:
   [1.71472, -0.0879841, -0.369562, -4.77127, 1.44898, -2.30881, 2.11664, -1.49164, 0.558957, -0.186297]

2. Text: "I have just adopted a cute dog"
   Fixed Code:
   [-0.5668864, -1.1441603, -2.3260381, -6.598249, 1.3243433, -1.7672672, 0.9366179, 0.3414038, 1.0347223, 1.2812055]
   LLaMA.cpp:
   [1.19416, -1.06577, -1.59795, -4.10442, 1.06202, -2.01701, 0.804208, 0.264081, 2.19793, 1.59068]

3. Text: "My favourite animal is the cat"
   Fixed Code:
   [2.5065503, 0.066849045, -0.081686154, -4.7165327, 1.6428406, -2.7072144, 2.914644, -1.7767899, 0.41250193, 0.5863753]
   LLaMA.cpp:
   [2.49985, 0.0733547, -0.093324, -4.74914, 1.63704, -2.69348, 2.94915, -1.73889, 0.421222, 0.616791]

[philpax]

Thank you for this! It's great to have this sorted out. The discrepancies between our numbers and llama.cpp are likely due to #210, which we should be able to resolve.

I revised your example to make it possible to use your own model / query / comparands, and to simplify all of the examples to reduce the amount of code to the bare essentials, but the spirit is more or less the same.

Again, thank you for addressing this 🙏