Add Talkie model support

[ZimengXiong]

Adds support for the Talkie language-model architecture and tokenizer so converted Talkie checkpoints can be loaded with mlx_lm.load / mlx_lm.generate.

Talkie is a GPT style model:

• RoPE with base 1,000,000
• Q/K RMS normalization after RoPE
• per-head attention gain
• per-block attention/MLP activation gains
• embedding skip connections
• untied LM head plus LM-head gain
• a tiktoken vocab.txt tokenizer with Talkie chat special tokens

PR adds:

• mlx_lm.models.talkie
• Talkie vocab.txt tokenizer adapter in tokenizer_utils.py
• Talkie chat template support
• tiktoken dependency for the tokenizer adapter

Validation

locally on M4 Max with a converted BF16 Talkie 1930 13B IT checkpoint.

PYTHONPATH=source/mlx-lm python - <<PY
from mlx_lm import load, generate
model, tokenizer = load("mlx/talkie-1930-13b-it-mlx")
prompt = tokenizer.apply_chat_template(
    [{"role":"user","content":"Write one sentence about radio."}],
    tokenize=False,
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=16, verbose=False))
PY

Output:

Radio was first successfully used for purposes of communication in 1912.

also,

PYTHONPATH=source/mlx-lm python -m mlx_lm.generate \
  --model mlx/talkie-1930-13b-it-mlx \
  --prompt "<|user|>Name a popular invention of the 1920s.<|end|><|assistant|>" \
  --max-tokens 12 \
  --temp 0

mlx weights at https://huggingface.co/zimengxiong/talkie-1930-13b-it-mlx .