gguf: use an unordered_map for faster duplicate tensor name lookups

[struct]

The current implementation iterates the vector of tensors comparing each name using strcmp. This operation gets slower with each new tensor added. We can improve the performance of this using an std::unordered_map which has a better average lookup time as the tensor vector grows in size.

Tested with:

$ build/bin/llama-bench 
| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| llama 1B Q4_0                  | 606.54 MiB |     1.10 B | Metal,BLAS |      12 |           pp512 |       5694.21 ± 8.22 |
| llama 1B Q4_0                  | 606.54 MiB |     1.10 B | Metal,BLAS |      12 |           tg128 |        339.54 ± 4.73 |

build: 38d3af1b (5955)

$ build/bin/llama-gguf models/7B/ggml-model-q4_0.gguf r n

$ build/bin/test-gguf 
...
108/108 tests passed
OK

 bash ./ci/run.sh ./tmp/results ./tmp/mnt

[struct]

This is a trivial change involving a temporary local unordered_map. I can't help but wonder why gguf_context doesn't hold similar data structures to improve lookups in gguf_get_key or gguf_get_tensor? Is there an explicit preference for these simple O(N) for loops?

[CISC]

Remnant from earlier iteration?

[CISC]

Suggested change

	auto [it, result] = tensor_names.emplace(info.t.name, i);
	auto [it, result] = tensor_names.emplace(info.t.name, i);