gguf-py: gguf_writer: Use bytearray to build metadata

[KerfuffleV2]

Repeatedly concatenating a bytes object to build the metadata in the GGUF writer is insanely slow. The part that mainly hurts is the vocab arrays, the non-array values are generally too small for the inefficiency to really be noticeable.

With this change, building the metadata for a model goes from around 5-6 sec to instant.

[cebtenzzre]

I don't think we need a file object here. Why not use bytearray? It has .append and .extend methods.

[KerfuffleV2]

Why not use bytearray?

The stuff I found seemed to indicate BytesIO was faster for this type of use case (appending chunks). This makes sense, since a bytearray has to efficiently support random access to individual elements. I would guess it's not a big difference though and I didn't benchmark it myself.

1. Poor performance when opening raw data with (wrong) gzip encoding in header mhe/pynrrd#88 (comment)
2. https://bugs.python.org/issue19801 (pretty old)

[KerfuffleV2]

@cebtenzzre Okay, so I tested it and they're exactly the same speed. I hacked convert.py to add the vocab metadata 10 times for a model with 64,000 vocab entries. It was like 3.020 sec for one and 3.019 for the other. Since using bytearray requires less modifications I switched to that.