ADR: StoryLM-0 loss operations

Context

storylm0_forward.kina emits [B, T, V] logits. Phase 1 gather cannot select a class from that rank-3 tensor. log(softmax(x)) is the wrong numerical form for a training loss.

Decision

Add two primitives, specified in 09-log-softmax-and-take-along-axis.md:

• log_softmax — same typing as softmax; stable max-shift + log(sum(exp)).
• take_along_axis — pick one element along a compile-time axis; indices have the values shape with that axis removed.

NLL is the composition -log_softmax.take_along_axis. No third primitive. AD VJPs are a later decision (36-storylm0-loss-ad.md).

Consequences

• The bootstrap .kina surface and the S-expression core both grow two forms.
• Host MLIR goldens and negative type tests land with the ops.
• Device take_along_axis and AD parity remain follow-on work.