fix(gemma3): make value_states contiguous before the KV cache append

[sidd-27]

gemma3::Model::forward fails for every input with slice-set only supports contiguous tensors.

value_states is transposed and then handed straight to KvCache::append, which slice_sets it into the cache buffer and requires a contiguous source. key_states gets away with it only incidentally — apply_rotary_emb_qkv calls .contiguous() on it first (gemma3.rs:109-110) — and value_states never goes through that path.

Reproduced on 0.10.2 and on current main. No downloads needed:

use candle_core::{DType, Device, Tensor};
use candle_nn::{VarBuilder, VarMap};
use candle_transformers::models::gemma3::{Config, Model};

fn cfg() -> Config {
    Config {
        attention_bias: false, head_dim: 8,
        hidden_activation: candle_nn::Activation::GeluPytorchTanh,
        hidden_size: 32, intermediate_size: 64,
        num_attention_heads: 4, num_hidden_layers: 2, num_key_value_heads: 2,
        rms_norm_eps: 1e-6, rope_theta: 10000.0, rope_local_base_freq: 10000.0,
        vocab_size: 64, final_logit_softcapping: None, attn_logit_softcapping: None,
        query_pre_attn_scalar: 8, sliding_window: 16, sliding_window_pattern: 2,
        max_position_embeddings: 128,
    }
}

fn main() {
    let device = Device::Cpu;
    let varmap = VarMap::new();
    let vb = VarBuilder::from_varmap(&varmap, DType::F32, &device);
    let mut model = Model::new(false, &cfg(), vb).unwrap();
    let ids = Tensor::new(&[[3u32, 9, 14, 2]], &device).unwrap();
    match model.forward(&ids, 0) {
        Ok(t) => println!("forward OK, shape {:?}", t.dims()),
        Err(e) => println!("forward FAILED: {e}"),
    }
}

Before:

forward FAILED: slice-set only supports contiguous tensors

After:

forward OK, shape [1, 1, 64]

It fails with sliding_window_pattern: 1 too, so both the Normal and Rotating cache paths are affected.

I suspect this has gone unnoticed because the recent Gemma 3 activity (#3745, #3747, #3326, #3299) is all against quantized_gemma3.rs, which has its own attention path.