feat: xlstm model and examples

[danielclough]

Add xLSTM (Extended LSTM) Model and Example

Adds support for xLSTM (Extended Long Short-Term Memory), a modernized LSTM architecture achieving competitive performance with Transformers while maintaining linear complexity for inference.

Implementation

Model architecture (candle-transformers/src/models/xlstm.rs):

• mLSTM blocks with matrix memory and exponential gating
• Covariance update rule using outer product of key-value pairs
• Stabilized gates with soft-capping and log-space computation
• GroupNorm without bias for multihead normalization
• SwiGLU FFN blocks with pre-norm residual connections

Text generation example (candle-examples/examples/xlstm/):

• Single-token recurrent inference with stateful generation
• Supports NX-AI/xLSTM-7b (~14GB VRAM in bf16, ~28GB in f32)
• Configurable sampling (temperature, top-p, repeat penalty)
• BOS token handling per model requirements

Usage

# Generate with default prompt (bf16, requires ~14GB VRAM)
cargo run --example xlstm --release --features cuda -- --prompt "Once upon a time" -n 50

# Use f32 precision
cargo run --example xlstm --release --features metal,accelerate -- --dtype f32 --prompt "The meaning of life is"