Fix multi-shard moe-split tensor counting

[i386]

This fixes llama-moe-split on the current master branch.

Users building Mesh-LLM's forked llama.cpp can now compile the MoE split tool again instead of failing with an undeclared n_tensors error in the multi-shard write path.

Root Cause

write_group() was updated to operate on input_shards, but the context-size calculation still referenced a single n_tensors variable from the older single-input implementation.

That made the current multi-shard path fail to compile on master.

Change

• sum gguf_get_n_tensors(...) across all input_shards
• size the output ggml context from that total tensor count

Validation

• cmake --build build --target llama-moe-split -j4