fix: restore MiniCPM inference after Granite Four changes

[jk3456a]

This commit fixes MiniCPM model inference that was broken by the Granite Four PR (#13550). The issue had two parts:

1. Missing LLM_KV_ATTENTION_LAYER_INDICES enum value that was removed, causing enum ordering to shift and breaking model metadata parsing

2. MiniCPM architecture uses llm_build_granite which was changed to use hparams.rope_finetuned instead of use_rope parameter, but MiniCPM models were not setting this flag correctly

Changes:

• Restore LLM_KV_ATTENTION_LAYER_INDICES enum and string mapping
• Set hparams.rope_finetuned = true for MiniCPM architecture

Fixes inference output from gibberish to correct model responses.

Tested with MiniCPM 0.5B model showing proper inference: Input: "你好"
Output: "你好，我是MiniCPM系列模型，由面壁智能和OpenBMB开源社区开发。详细信息请访问 https://github.com/OpenBMB/ [end of text]"

Make sure to read the contributing guidelines before submitting a PR

[ggerganov]

Missing LLM_KV_ATTENTION_LAYER_INDICES enum value that was removed, causing enum ordering to shift and breaking model metadata parsing

Hm, I don't think the actual enum values matter. Which parsing is broken?

[jk3456a]

sorry you are right, the enum value doesn't matter.
The actual issue was that MiniCPM models use the llm_build_granite constructor, which was changed in the Granite Four PR to use hparams.rope_finetuned instead of the use_rope parameter. MiniCPM models need rope enabled by default, but weren't setting this flag.