feat(gemma3): add embed_tokens() and forward_embeds()

[sidd-27]

gemma3::Model only accepts token ids, and keeps embed_tokens private. A multimodal caller needs to build a mixed sequence — text embeddings with image embeddings spliced in — which has no representation as token ids, so Gemma 3's vision variants can't be wired up from outside the crate.

gemma.rs already exposes both entry points (embed_tokens() at line 371, forward_embeds() at line 412) and paligemma.rs is built on them. This adds the same pair to gemma3.

forward now delegates to forward_embeds, so there's no behaviour change. Unlike gemma, forward_embeds builds the attention masks itself rather than taking them as an argument, because gemma3 uses two (full and sliding) and picks per layer — having the caller supply them would leak that detail.

Verified the two paths agree exactly (tiny randomly-initialised model, CPU, no downloads):

let a = model.forward(&ids, 0).unwrap();
model.clear_kv_cache();
let embeds = model.embed_tokens().forward(&ids).unwrap();
let b = model.forward_embeds(&embeds, 0).unwrap();

logits len              = 64
max |forward - embeds|  = 0.000000000
OK: forward and forward_embeds are identical

Note that running that check at all requires #3902 — gemma3::Model::forward currently fails for every input with slice-set only supports contiguous tensors. This PR compiles and is independent of it, but can't be exercised at runtime until that one lands.