Add Llama callPartial for pipeline-parallel inference

[anupsv]

Summary

Adds callPartial entry points on LlamaModelInner and LlamaModel so a single rank can run a contiguous sub-range of transformer blocks, with control over whether the rank applies token embedding, the final RMS norm, and the LM head.

Used by downstream pipeline-parallel inference (e.g. d-inference EncryptedPipelineInference.swift / PipelineInference.swift) where two Macs split a Llama-family model across Thunderbolt 5 and exchange activation tensors between ranks.

Shape of the change

Method	Where	Purpose
LlamaModelInner.callPartial(_:layerRange:applyEmbedding:applyNorm:cache:)	Libraries/MLXLLM/Models/Llama.swift	Run blocks layerRange.lowerBound ..< layerRange.upperBound. Rank 0 sets applyEmbedding=true; last rank sets applyNorm=true. KV cache indices are relative to layerRange.
LlamaModel.callPartial(_:layerRange:applyEmbedding:applyNorm:applyHead:cache:)	Libraries/MLXLLM/Models/Llama.swift	Wraps LlamaModelInner.callPartial and optionally projects to vocab logits via the LM head (last rank only).

Net: 1 file, +52 lines, additive. No existing forward path changes.

Test plan

• swift build succeeds on macOS 14+ (Apple Silicon)
• Existing Llama / Mistral consumers still compile and produce identical outputs (verified via d-inference provider-swift tests)
• Two-rank pipeline split on a Llama-class model produces logits equal to a single-rank run (within numerical tolerance)

Why upstream

Pinning this on our fork unblocks d-inference CI (the d-inference encrypted-pipeline-inference stack imports callPartial directly). Once landed we'll bump the d-inference submodule pointer. Related: Layr-Labs/mlx-swift#2 (the matching change on the underlying mlx-swift fork to expose Cmlx + enable jaccl).