fix #149 - load tensors by type, ignoring filetype

[philpax]

As noticed by @KerfuffleV2, both loaders incorrectly handle newer models that mix tensor types indiscriminately due to a confusion between the ftype/f16_ and element type.

I was planning on bringing #84 up to date first, but realised that I needed to figure out what was going on with ftype before baking that in.

I've addressed the issue by decoupling them properly, and then switching over to loading the tensors from file instead of trying to preallocate them with the wrong type. This mirrors the changes made in ggerganov/llama.cpp#801.

I've tested this with Alpaca 7B GGML and gpt4-x-alpaca-13b-native-4bit-128g, the latter with and without mmap, and all seems to work.

---

I'm not happy at how I broke the isolation between the loader and the Model with Model::new_loader2, but I'm going to revisit this once I remove loader1 as part of #150.

loader1 is still broken (i.e. using the old behaviour), but that's OK because it's going away soon ^_^