feat: allow tokenizer to load from GGUF metadata

[fffonion]

Right now a subset of pipeline are supported (qwen2, llama3 style BPE, gpt2).

Example:

cargo run --example gguf-tokenizer -- --model unsloth/Qwen3-4B-GGUF/Qwen3-4B-Q4_K_S.gguf --prompt "Hello"

To use in your code:

use candle_core::quantized::tokenizer::TokenizerFromGguf;

let content = gguf_file::Content::read(&mut reader).context("failed to load GGUF metadata")?;
let tokenizer =
        Tokenizer::from_gguf(&content).context("failed to initialize tokenizer from GGUF")?;