Non-record: LeakyMixer: 11L leaky_relu(0.5)^2 + backoff n-gram mixer

[zlxi02]

• swapped relu^2 for leaky_relu(0.5)^2, bumped to 11 layers, and added a warmdown of 3500 steps. also wrote a backoff n-gram mixer in C that runs at eval time; builds a token cache as it scores the val set and mixes with the neural logits using entropy-adaptive alpha.
• ran on 4xA100 SXM (was low on credits) so only got 849 steps. val_bpb came out to 1.4096 neural-only. on 8xH100 this should get ~1700 steps and around 1.20 BPB, with the n-gram mixer dropping it further.
• artifact size: 13.49 MB