Topk annealing

[adamkarvonen]

3 changes here:

• First, add an optional k-annealing schedule for TopK trainers. With this, K begins at d_model and is annealed to K over the first 10% of training. Multiple people have found a significant reduction in dead features, such as Llama Scope: https://arxiv.org/abs/2410.20526
• Add Qwen3 submodule
• Handle the remove_bos argument if the tokenizer does not contain a BOS token (such as Qwen3). In this case, we instead remove the first non-pad token of each sequence, as it still contains a high norm. For Qwen3-8B, the first non-pad token typically has a norm 100x the average.