llama.cpp Adds SarvamMoE Support

llama.cpp PR #20275 adds a dedicated sarvam_moe path.

• Extends the structure for Sarvam-30B based on BailingMoeForCausalLM.
• 19-layer configuration: 1 dense FFN + 18 MoE layers, 128 routed experts, top-6, 1 shared expert.
• Reflects sigmoid gating, zero-mean expert bias normalization, and routed_scaling_factor=2.5.
• Also includes GQA (64 heads, 4 KV heads), combined QKV projection, and QK RMSNorm.
• The tokenizer handles converting SentencePiece BPE to GPT-2 byte-level encoding.

While the official Sarvam GGUF can also be loaded via the existing bailingmoe2 path, this PR focuses on adding a dedicated Sarvam architecture and conversion logic to correctly align RoPE, expert bias, and tokenizer handling. Testing confirmed HF safetensors conversion, multiple quantizations (Q8_0/Q6_K/Q4_K_M), and inference behavior. However, a comment on May 8th pointed out issues with Indic script and English tokenizer consistency, and the maintainer requested a rebase onto master.