Recursive Latent Forcing ports to GPT-2

This is the ablation that matters: if the same scaffold works on GPT-2, RLF starts to look like a portable training recipe rather than a Mamba-specific hack.

• –The frozen GPT-2 pass plus the 14M loop core keeps the compute story compelling, because the heavy backbone runs once and repeated reasoning stays cheap.
• –The 6-hop and 10-hop wins are real, but the 7-hop `saxophone` miss is a tokenizer artifact, not a clean reasoning victory.
• –The 8-hop early halt suggests the method is still brittle at the edge of its depth budget, and the slight gap versus the original Mamba2 run hints that backbone choice still matters.
• –The key unanswered test is whether GPT-2 can drop the lifeline at inference the way Mamba2 did; if it can, this becomes a general recipe for implicit multi-step reasoning without chain-of-thought tokens.